TFRS_Demo.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "035a4fe5",
   "metadata": {},
   "source": [
    "# TensorFlow Recommenders (TFRS) Demo \n",
    "\n",
    "In this notebook we will demonstrate the use of TFRS on sales data from [Babyshop](https://www.babyshop.se/). This tutorial is heavily based on the [official tensorflow recommenders tutorials](https://www.tensorflow.org/recommenders/examples/quickstart). \n",
    "\n",
    "<div class=\"alert alert-block alert-info\">\n",
    "There are a lot of machine learning \"best practices\" that are ignored in this notebook for the sake of simplicity. The focus is to get an introduction to TFRS and general understanding of how this library works, not to build an industrial recommendation system. \n",
    "</div>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1f6cf9a3",
   "metadata": {},
   "source": [
    "**YOU WILL NEED TO RUN THE FOLLOWING CELL IF USING GOOGLE COLAB**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "16f6b2c6",
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install tensorflow_recommenders==0.6.0"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a202403a",
   "metadata": {},
   "source": [
    "**AFTER YOU INSTALL TENSORFLOW RECOMMENDERS YOU MAY NEED TO RESTART THE RUNTIME** \n",
    "\n",
    "\"Runtime\" --> \"Restart Runtime\""
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7b6d1473",
   "metadata": {},
   "source": [
    "## **Imports**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "12c891d9",
   "metadata": {},
   "outputs": [],
   "source": [
    "from typing import Dict, Any, Text\n",
    "\n",
    "import numpy as np \n",
    "import pandas as pd\n",
    "\n",
    "import tensorflow as tf\n",
    "import tensorflow_recommenders as tfrs"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "61f4d4a7",
   "metadata": {},
   "source": [
    "# **Reading in the Data** \n",
    "\n",
    "First we will read in the training and test data. \n",
    "\n",
    "<div class=\"alert alert-block alert-info\">\n",
    "See <code>EDA.ipynb</code> for analysis on the data and details on how the train and test sets were created. \n",
    "</div>\n",
    "\n",
    "In the following cells we will create an even smaller version of the dataset so that we can train on a reasonable amount of time on a CPU. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "2fbc8ca6",
   "metadata": {},
   "outputs": [],
   "source": [
    "TRAIN_PATH = 'https://raw.githubusercontent.com/msvensson222/tfrs-retail-example/master/train.csv'\n",
    "TEST_PATH = 'https://raw.githubusercontent.com/msvensson222/tfrs-retail-example/master/test.csv'\n",
    "ITEM_INFO_PATH = 'https://raw.githubusercontent.com/msvensson222/tfrs-retail-example/master/item_info.csv'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "f81fd908",
   "metadata": {},
   "outputs": [],
   "source": [
    "train_df = pd.read_csv(TRAIN_PATH, dtype={'user_no': str, 'item_no': str})\n",
    "test_df = pd.read_csv(TEST_PATH, dtype={'user_no': str, 'item_no': str})\n",
    "\n",
    "# For evaluation\n",
    "item_info_df = pd.read_csv(ITEM_INFO_PATH, dtype={'item_no': str})"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "11a551c2",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>user_no</th>\n",
       "      <th>item_no</th>\n",
       "      <th>gender_description</th>\n",
       "      <th>brand</th>\n",
       "      <th>product_group</th>\n",
       "      <th>first_interaction_month</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>3514657341026450752</td>\n",
       "      <td>-8200171396217105230</td>\n",
       "      <td>girls</td>\n",
       "      <td>jacadi</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>5</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>-2544835772752526495</td>\n",
       "      <td>6010486836306001722</td>\n",
       "      <td>unisex</td>\n",
       "      <td>done by deer</td>\n",
       "      <td>tableware</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>-6023760384625599940</td>\n",
       "      <td>-289310928076258010</td>\n",
       "      <td>unisex</td>\n",
       "      <td>axkid</td>\n",
       "      <td>car seat accessories</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>4084143572023326121</td>\n",
       "      <td>-1069008842172275553</td>\n",
       "      <td>boys</td>\n",
       "      <td>pom dapi</td>\n",
       "      <td>sandals</td>\n",
       "      <td>5</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>-4787976733877481713</td>\n",
       "      <td>608763176274829755</td>\n",
       "      <td>unisex</td>\n",
       "      <td>little luwi</td>\n",
       "      <td>tops</td>\n",
       "      <td>8</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>667004</th>\n",
       "      <td>6183491195824661353</td>\n",
       "      <td>-487489333946043722</td>\n",
       "      <td>girls</td>\n",
       "      <td>ikks</td>\n",
       "      <td>dresses</td>\n",
       "      <td>4</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>667005</th>\n",
       "      <td>-8074445800271606192</td>\n",
       "      <td>7154496603299236573</td>\n",
       "      <td>unisex</td>\n",
       "      <td>tommee tippee</td>\n",
       "      <td>baby feeding</td>\n",
       "      <td>3</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>667006</th>\n",
       "      <td>3873852775369901008</td>\n",
       "      <td>3465194094158419708</td>\n",
       "      <td>unisex</td>\n",
       "      <td>by nils</td>\n",
       "      <td>sandals</td>\n",
       "      <td>5</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>667007</th>\n",
       "      <td>-1306455725574612144</td>\n",
       "      <td>2424760068735106973</td>\n",
       "      <td>girls</td>\n",
       "      <td>kenzo</td>\n",
       "      <td>tops</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>667008</th>\n",
       "      <td>3762176132479070423</td>\n",
       "      <td>2648857732393394528</td>\n",
       "      <td>unisex</td>\n",
       "      <td>kids concept</td>\n",
       "      <td>role play</td>\n",
       "      <td>9</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>667009 rows × 6 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "                     user_no               item_no gender_description  \\\n",
       "0        3514657341026450752  -8200171396217105230              girls   \n",
       "1       -2544835772752526495   6010486836306001722             unisex   \n",
       "2       -6023760384625599940   -289310928076258010             unisex   \n",
       "3        4084143572023326121  -1069008842172275553               boys   \n",
       "4       -4787976733877481713    608763176274829755             unisex   \n",
       "...                      ...                   ...                ...   \n",
       "667004   6183491195824661353   -487489333946043722              girls   \n",
       "667005  -8074445800271606192   7154496603299236573             unisex   \n",
       "667006   3873852775369901008   3465194094158419708             unisex   \n",
       "667007  -1306455725574612144   2424760068735106973              girls   \n",
       "667008   3762176132479070423   2648857732393394528             unisex   \n",
       "\n",
       "                brand         product_group  first_interaction_month  \n",
       "0              jacadi           all in ones                        5  \n",
       "1        done by deer             tableware                       11  \n",
       "2               axkid  car seat accessories                        2  \n",
       "3            pom dapi               sandals                        5  \n",
       "4         little luwi                  tops                        8  \n",
       "...               ...                   ...                      ...  \n",
       "667004           ikks               dresses                        4  \n",
       "667005  tommee tippee          baby feeding                        3  \n",
       "667006        by nils               sandals                        5  \n",
       "667007          kenzo                  tops                        1  \n",
       "667008   kids concept             role play                        9  \n",
       "\n",
       "[667009 rows x 6 columns]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "display(train_df)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8f742982",
   "metadata": {},
   "source": [
    "We will obtain our smaller dataset by just taking the top 2000 users (i.e. users with the most interactions) in the training data. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "8c64cee9",
   "metadata": {},
   "outputs": [],
   "source": [
    "NUM_USERS = 2000\n",
    "top_users = train_df['user_no'].value_counts()[:NUM_USERS].index\n",
    "\n",
    "train_df_filtered = train_df.loc[train_df['user_no'].isin(top_users), :]\n",
    "test_df_filtered = test_df.loc[test_df['user_no'].isin(top_users), :]\n",
    "items = train_df_filtered['item_no'].unique()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d260ea11",
   "metadata": {},
   "source": [
    "In the following cells we create TensorFlow datasets out of the Pandas DataFrames and print out the first few instances just to get an idea of what the datasets look like. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "12c54f0e",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "2022-02-10 09:28:28.824445: I tensorflow/core/platform/cpu_feature_guard.cc:151] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 FMA\n",
      "To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.\n"
     ]
    }
   ],
   "source": [
    "train_dataset = tf.data.Dataset.from_tensor_slices(dict(train_df_filtered))\n",
    "test_dataset = tf.data.Dataset.from_tensor_slices(dict(test_df_filtered))\n",
    "\n",
    "items_dataset = tf.data.Dataset.from_tensor_slices(items)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "82c9ae58",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tf.Tensor(b'-1119687312509640915', shape=(), dtype=string)\n",
      "tf.Tensor(b'-3219910350938683317', shape=(), dtype=string)\n",
      "tf.Tensor(b'1179978263120783371', shape=(), dtype=string)\n"
     ]
    }
   ],
   "source": [
    "for item in items_dataset.take(3):\n",
    "    print(item)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "0a3ab3d4",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'user_no': <tf.Tensor: shape=(), dtype=string, numpy=b'-2683506524939646253'>, 'item_no': <tf.Tensor: shape=(), dtype=string, numpy=b'-1119687312509640915'>, 'gender_description': <tf.Tensor: shape=(), dtype=string, numpy=b'unisex'>, 'brand': <tf.Tensor: shape=(), dtype=string, numpy=b'reima'>, 'product_group': <tf.Tensor: shape=(), dtype=string, numpy=b'boots'>, 'first_interaction_month': <tf.Tensor: shape=(), dtype=int64, numpy=11>}\n",
      "{'user_no': <tf.Tensor: shape=(), dtype=string, numpy=b'-8270295623916047084'>, 'item_no': <tf.Tensor: shape=(), dtype=string, numpy=b'-3219910350938683317'>, 'gender_description': <tf.Tensor: shape=(), dtype=string, numpy=b'boys'>, 'brand': <tf.Tensor: shape=(), dtype=string, numpy=b'moschino kid-teen'>, 'product_group': <tf.Tensor: shape=(), dtype=string, numpy=b'tops'>, 'first_interaction_month': <tf.Tensor: shape=(), dtype=int64, numpy=11>}\n",
      "{'user_no': <tf.Tensor: shape=(), dtype=string, numpy=b'-1493854771764820101'>, 'item_no': <tf.Tensor: shape=(), dtype=string, numpy=b'1179978263120783371'>, 'gender_description': <tf.Tensor: shape=(), dtype=string, numpy=b'girls'>, 'brand': <tf.Tensor: shape=(), dtype=string, numpy=b'molo'>, 'product_group': <tf.Tensor: shape=(), dtype=string, numpy=b'tops'>, 'first_interaction_month': <tf.Tensor: shape=(), dtype=int64, numpy=9>}\n"
     ]
    }
   ],
   "source": [
    "for interaction in train_dataset.take(3):\n",
    "    print(interaction)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3031ff63",
   "metadata": {},
   "source": [
    "---\n",
    "---\n",
    "\n",
    "# **Baseline**\n",
    "\n",
    "The first thing we can do is start with a very \"naive\" baseline: for every interaction in the test dataset we will just predict the top 100 items from the training set. This will give us a reference point for any metrics we calculate after training a model. \n",
    "\n",
    "A side benefit is that we can get a better understanding of TFRS by recreating the way that metrics are calculated by TFRS. See [here](https://github.com/tensorflow/recommenders/blob/8b249f3fc0f8d3d907eecf010809a5df3759d65d/tensorflow_recommenders/metrics/factorized_top_k.py#L64) for the source code; the following cells are basically a simplified version of the code found in the TFRS library. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "0d9ac392",
   "metadata": {},
   "outputs": [],
   "source": [
    "NUM_TOP_ITEMS = 100\n",
    "top_items = train_df_filtered['item_no'].value_counts()[:100].index\n",
    "\n",
    "ks = (1, 5, 10, 50, 100)\n",
    "metrics = [tf.keras.metrics.Mean() for k in ks]\n",
    "\n",
    "true_candidates = tf.expand_dims(tf.constant(test_df_filtered['item_no'].values), 1)\n",
    "retrieved_candidates = tf.expand_dims(top_items, 1)\n",
    "# Pretend like we retrieve the same top 100 candidates for every interaction in test data\n",
    "retrieved_candidates = tf.transpose(tf.repeat(retrieved_candidates, \n",
    "                                              tf.constant(true_candidates.shape[0]), \n",
    "                                              axis=1))\n",
    "ids_match = tf.cast(tf.math.equal(true_candidates, retrieved_candidates), tf.float32)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "0500e11c",
   "metadata": {},
   "outputs": [],
   "source": [
    "for k, metric in zip(ks, metrics):\n",
    "    # Clip to only count multiple matches once.\n",
    "    match_found = tf.clip_by_value(\n",
    "        tf.reduce_sum(ids_match[:, :k], axis=1, keepdims=True),\n",
    "        0.0, 1.0\n",
    "    )\n",
    "    metric.update_state(match_found)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "5162a868",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Top 1 categorical accuracy: 0.00218\n",
      "Top 5 categorical accuracy: 0.00480\n",
      "Top 10 categorical accuracy: 0.01134\n",
      "Top 50 categorical accuracy: 0.02967\n",
      "Top 100 categorical accuracy: 0.04058\n"
     ]
    }
   ],
   "source": [
    "for k, metric in zip(ks, metrics):\n",
    "    print(f'Top {k} categorical accuracy: {metric.result().numpy():.5f}')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "65826210",
   "metadata": {},
   "source": [
    "# Creating a Simple Model\n",
    "\n",
    "We will start by creating a very simple model similar to the one created in [the TFRS basic retrieval tutorial](https://www.tensorflow.org/recommenders/examples/basic_retrieval). Quoting from the tutorial, the model will be created by composing two sub-models: \n",
    "\n",
    "> 1. A query model computing the query representation (normally a fixed-dimensionality embedding vector) using query features\n",
    "> 2. A candidate model computing the candidate representation (an equally-sized vector using the candidate features\n",
    "> \n",
    "> The outputs of the two models are then multiplied together to give a query-candidate affinity score, with higher scores expressing a better match between the candidate and the query.\n",
    "\n",
    "For our use case, we will pretend that we want to recommend items to users. As such, our **query** model will produce representations of the **users** (and potentially additional **context**, such as time, device, etc.) and our **candidate** model will produce representations of the **items**. \n",
    "\n",
    "For the rest of the notebook we will refer to the \"query\" model as a `user_model` and the \"candidate\" model as a `item_model`\n",
    "\n",
    "<div class=\"alert alert-block alert-info\">\n",
    "<b>Tip:</b>  There is nothing forcing us to associate users with a query model and items with a candidate model. For example, we could just as easily associate items with a query model and items with a candidate model for an <b>item-item</b> recommender. \n",
    "</div>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ae1a9a04",
   "metadata": {},
   "source": [
    "In the following cells we will build each tower separately (via the `create_embedding_model` function). We will also define the task, which in this case will be a retrieval task. Finally we will put together the two sub-models and the task in a `tfrs.Model`, which allows us to implement a model by only implementing the `__init__` and `compute_loss` methods—the base model class will take care of the training loop. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "7a734989",
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_vocab(df, feature, top_n=None):\n",
    "    return df[feature].value_counts()[:top_n].index\n",
    "\n",
    "def create_embedding_model(df, feature, num_oov_indices=1, embedding_dim=32):\n",
    "    feature_vocab = get_vocab(df, feature)\n",
    "    embedding_model = tf.keras.Sequential([\n",
    "        tf.keras.layers.StringLookup(vocabulary=feature_vocab, \n",
    "                                     num_oov_indices=num_oov_indices),\n",
    "        tf.keras.layers.Embedding(len(feature_vocab) + num_oov_indices, embedding_dim)\n",
    "    ])\n",
    "    \n",
    "    \n",
    "    return embedding_model\n",
    "\n",
    "class SimpleTFRSModel(tfrs.Model):\n",
    "\n",
    "    def __init__(self, user_model, item_model, task):\n",
    "        super().__init__()\n",
    "        self.user_model: tf.keras.Model = user_model\n",
    "        self.item_model: tf.keras.Model = item_model\n",
    "        self.task: tf.keras.layers.Layer = task\n",
    "            \n",
    "\n",
    "    def compute_loss(self, features: Dict[Text, tf.Tensor], training=False) -> tf.Tensor:\n",
    "        # We pick out the user features and pass them into the user model\n",
    "        # and item features to pass to the item model. Use the returned embeddings \n",
    "        # to calculate the loss\n",
    "        user_embeddings = self.user_model(features['user_no'])\n",
    "        positive_item_embeddings = self.item_model(features['item_no'])\n",
    "        # The task computes the loss and the metrics. Don't compute metrics during training \n",
    "        # because it will take too long otherwise\n",
    "        return self.task(user_embeddings, positive_item_embeddings, compute_metrics=not training)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "a70e26bb",
   "metadata": {},
   "outputs": [],
   "source": [
    "user_model = create_embedding_model(train_df_filtered, \"user_no\")\n",
    "item_model = create_embedding_model(train_df_filtered, \"item_no\")\n",
    "metrics = tfrs.metrics.FactorizedTopK(\n",
    "  candidates=items_dataset.batch(128).map(item_model)\n",
    ")\n",
    "task = tfrs.tasks.Retrieval(\n",
    "  metrics=metrics\n",
    ")\n",
    "\n",
    "simple_tfrs_model = SimpleTFRSModel(user_model, item_model, task)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7778cb09",
   "metadata": {},
   "source": [
    "---\n",
    "---\n",
    "\n",
    "<div class=\"alert alert-block alert-warning\">\n",
    "<b>The above is just a convenience!</b> The following class is a simplified version of what\n",
    "is actually going on under-the-hood:\n",
    "\n",
    "```python \n",
    "class NonTFRSModel(tf.keras.Model):\n",
    "    def __init__(self, user_model, item_model, metrics):\n",
    "        \"\"\"\n",
    "        Note that we don't pass in the task! That's because we define \n",
    "        what the task is here.\n",
    "        \"\"\"\n",
    "        super().__init__()\n",
    "        self.user_model = user_model \n",
    "        self.item_model = item_model \n",
    "        # When we perform retrieval, the default loss is actually just good \n",
    "        # old CategoricalCrossentropy :) \n",
    "        self._loss = tf.keras.losses.CategoricalCrossentropy(\n",
    "            from_logits=True, reduction=tf.keras.losses.Reduction.SUM\n",
    "        )\n",
    "        self._factorized_metrics = metrics\n",
    "\n",
    "    def calc_loss(self, query_embeddings, candidate_embeddings): \n",
    "        scores = tf.linalg.matmul(\n",
    "            query_embeddings, \n",
    "            candidate_embeddings, \n",
    "            transpose_b=True\n",
    "        )\n",
    "        num_queries, num_candidates = scores.shape\n",
    "        labels = tf.eye(num_queries, num_candidates)\n",
    "        loss = self._loss(y_true=labels, y_pred=scores)\n",
    "        self._factorized_metrics.update_state(\n",
    "            query_embeddings, \n",
    "            candidate_embeddings\n",
    "        )\n",
    "        return loss\n",
    "    \n",
    "\n",
    "    def train_step(self, features: Dict[Text, tf.Tensor]) -> tf.Tensor:\n",
    "        with tf.GradientTape() as tape: \n",
    "            user_embeddings = self.user_model(features['user_no'])\n",
    "            positive_item_embeddings = self.item_model(features['item_no'])\n",
    "            loss = self.calc_loss(user_embeddings, positive_item_embeddings)\n",
    "\n",
    "        gradients = tape.gradient(loss, self.trainable_variables)\n",
    "        self.optimizer.apply_gradients(zip(gradients, self.trainable_variables))\n",
    "\n",
    "        metrics = {metric.name: metric.result() for metric in self.metrics}\n",
    "        return metrics \n",
    "\n",
    "    def test_step(self, features: Dict[Text, tf.Tensor]) -> tf.Tensor: \n",
    "        user_embeddings = self.user_model(features['user_no'])\n",
    "        positive_item_embeddings = self.item_model(features['item_no'])\n",
    "\n",
    "        loss = self.compute_loss(user_embeddings, positive_item_embeddings)        \n",
    "\n",
    "        metrics = {metric.name: metric.result() for metric in self.metrics}\n",
    "        return metrics \n",
    "```\n",
    "\n",
    "We can then instantiate and compile a model like so: \n",
    "\n",
    "```python \n",
    "simple_model = NonTFRSModel(user_model, item_model, metrics)\n",
    "# Need to specify run_eagerly=True because we need the shape of the scores \n",
    "# in the calc_loss function\n",
    "simple_model.compile(optimizer=tf.keras.optimizers.Adam(), run_eagerly=True)\n",
    "```\n",
    "\n",
    "After that we can just train the model the same as below :)\n",
    "\n",
    "</div>\n",
    "---\n",
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "b7811f34",
   "metadata": {},
   "outputs": [],
   "source": [
    "train_dataset_interactions = train_dataset.map(lambda x: {\n",
    "    'user_no': x['user_no'],\n",
    "    'item_no': x['item_no']\n",
    "})\n",
    "test_dataset_interactions = test_dataset.map(lambda x: {\n",
    "    'user_no': x['user_no'],\n",
    "    'item_no': x['item_no']\n",
    "})\n",
    "\n",
    "train_ds = train_dataset_interactions.shuffle(1_000).batch(4096)\n",
    "test_ds = test_dataset_interactions.batch(4096)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7839e0b4",
   "metadata": {},
   "source": [
    "<div class=\"alert alert-block alert-info\">\n",
    "In the interest of accelerating training as much as possible, we won't calculate any metrics. If we were training \"for real\" we'd probably want to monitor the training and implement early stopping \n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "f5c6a408",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/3\n",
      "10/10 [==============================] - 9s 795ms/step - factorized_top_k/top_1_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_5_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_10_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_50_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_100_categorical_accuracy: 0.0000e+00 - loss: 27906.1022 - regularization_loss: 0.0000e+00 - total_loss: 27906.1022 - val_factorized_top_k/top_1_categorical_accuracy: 0.0000e+00 - val_factorized_top_k/top_5_categorical_accuracy: 0.0017 - val_factorized_top_k/top_10_categorical_accuracy: 0.0031 - val_factorized_top_k/top_50_categorical_accuracy: 0.0092 - val_factorized_top_k/top_100_categorical_accuracy: 0.0118 - val_loss: 17733.2559 - val_regularization_loss: 0.0000e+00 - val_total_loss: 17733.2559\n",
      "Epoch 2/3\n",
      "10/10 [==============================] - 6s 668ms/step - factorized_top_k/top_1_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_5_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_10_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_50_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_100_categorical_accuracy: 0.0000e+00 - loss: 27767.1332 - regularization_loss: 0.0000e+00 - total_loss: 27767.1332 - val_factorized_top_k/top_1_categorical_accuracy: 4.3630e-04 - val_factorized_top_k/top_5_categorical_accuracy: 0.0057 - val_factorized_top_k/top_10_categorical_accuracy: 0.0087 - val_factorized_top_k/top_50_categorical_accuracy: 0.0253 - val_factorized_top_k/top_100_categorical_accuracy: 0.0358 - val_loss: 17709.3477 - val_regularization_loss: 0.0000e+00 - val_total_loss: 17709.3477\n",
      "Epoch 3/3\n",
      "10/10 [==============================] - 7s 686ms/step - factorized_top_k/top_1_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_5_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_10_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_50_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_100_categorical_accuracy: 0.0000e+00 - loss: 26974.9058 - regularization_loss: 0.0000e+00 - total_loss: 26974.9058 - val_factorized_top_k/top_1_categorical_accuracy: 4.3630e-04 - val_factorized_top_k/top_5_categorical_accuracy: 0.0057 - val_factorized_top_k/top_10_categorical_accuracy: 0.0096 - val_factorized_top_k/top_50_categorical_accuracy: 0.0332 - val_factorized_top_k/top_100_categorical_accuracy: 0.0506 - val_loss: 17616.6504 - val_regularization_loss: 0.0000e+00 - val_total_loss: 17616.6504\n"
     ]
    }
   ],
   "source": [
    "simple_tfrs_model.compile(optimizer=tf.keras.optimizers.Adagrad(0.1))\n",
    "history = simple_tfrs_model.fit(train_ds, \n",
    "                                epochs=3, \n",
    "                                validation_data=test_ds)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "18e3d1ff",
   "metadata": {},
   "source": [
    "## Evaluation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "c20ad450",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "10/10 [==============================] - 55s 5s/step - factorized_top_k/top_1_categorical_accuracy: 0.0251 - factorized_top_k/top_5_categorical_accuracy: 0.1320 - factorized_top_k/top_10_categorical_accuracy: 0.2429 - factorized_top_k/top_50_categorical_accuracy: 0.6843 - factorized_top_k/top_100_categorical_accuracy: 0.8510 - loss: 25361.1767 - regularization_loss: 0.0000e+00 - total_loss: 25361.1767\n",
      "1/1 [==============================] - 3s 3s/step - factorized_top_k/top_1_categorical_accuracy: 4.3630e-04 - factorized_top_k/top_5_categorical_accuracy: 0.0057 - factorized_top_k/top_10_categorical_accuracy: 0.0096 - factorized_top_k/top_50_categorical_accuracy: 0.0332 - factorized_top_k/top_100_categorical_accuracy: 0.0506 - loss: 17616.6504 - regularization_loss: 0.0000e+00 - total_loss: 17616.6504\n"
     ]
    }
   ],
   "source": [
    "train_results = simple_tfrs_model.evaluate(train_ds, return_dict=True)\n",
    "test_results = simple_tfrs_model.evaluate(test_ds, return_dict=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "d9516496",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train top-100 accuracy:  0.8510119318962097\n",
      "Test top-100 accuracy:  0.050610821694135666\n"
     ]
    }
   ],
   "source": [
    "print(f\"Train top-100 accuracy:  {train_results['factorized_top_k/top_100_categorical_accuracy']}\")\n",
    "print(f\"Test top-100 accuracy:  {test_results['factorized_top_k/top_100_categorical_accuracy']}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a214ec1e",
   "metadata": {},
   "source": [
    "The model does slightly better than the baseline, but it is also overfitting like crazy. Quoting from the tutorial, this is due to two factors: \n",
    "\n",
    "> 1. Our model is likely to perform better on the data that it has seen, simply because it can memorize it. This overfitting phenomenon is especially strong when models have many parameters. It can be mediated by model regularization and use of user and movie features that help the model generalize better to unseen data.\n",
    "> 2. The model is re-recommending some of users' already [bought items]. These known-positive watches can crowd out test [items] out of top K recommendations."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c128f4db",
   "metadata": {},
   "source": [
    "## Serving and Qualitative Evaluation"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "451d33e5",
   "metadata": {},
   "source": [
    "In order to serve the model, we create an \"index\". Basically this is a way for us to do nearest neighbor search in the embedding space: we get in a \"query\" (in this case a user), calculate an embedding, and then compare that embedding to the embeddings of all candidate items. \n",
    "\n",
    "In this case, the number of candidate items is very small, so we just brute force the search. For real-world use cases we would want to use an approximate nearest neighbor search. TFRS allows us to build an index based on [ScaNN](https://github.com/google-research/google-research/tree/master/scann) if we install the optional dependency. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "bfea60c5",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Create a model that takes in raw query features, and\n",
    "index = tfrs.layers.factorized_top_k.BruteForce(simple_tfrs_model.user_model)\n",
    "# recommends items out of the entire items dataset.\n",
    "_ = index.index_from_dataset(\n",
    "        tf.data.Dataset.zip((items_dataset.batch(100), \n",
    "                             items_dataset.batch(100).map(simple_tfrs_model.item_model))))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "121c8135",
   "metadata": {},
   "source": [
    "To qualitatively analyze the performance of the model, we can look at the predictions for a random user. \n",
    "\n",
    "<div class=\"alert alert-block alert-info\">\n",
    "<b>Tip: </b> Rerun the next few cells to get predictions for different users. \n",
    "    \n",
    "<b>Tip: </b> The first time you run the <code>%%time</code> cells, it may take a bit longer than normal if there is any <code>@tf.function</code> tracing going on. \n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "6960f930",
   "metadata": {},
   "outputs": [],
   "source": [
    "random_user = np.random.choice(train_df_filtered['user_no'].unique())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "a0938ee9",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 4.17 ms, sys: 2.01 ms, total: 6.18 ms\n",
      "Wall time: 11.2 ms\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "# Get recommendations.\n",
    "_, titles = index(tf.constant([random_user]))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3cc5af51",
   "metadata": {},
   "source": [
    "<div class=\"alert alert-block alert-info\">\n",
    "<b>Tip: </b> You can also explicitly exclude certain items (e.g. items prviously interacted with) by calling <code>query_with_exclusions</code>. \n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "76cdf34c",
   "metadata": {},
   "outputs": [],
   "source": [
    "items_to_exclude = train_df_filtered.loc[train_df_filtered['user_no'] == random_user]['item_no'].unique()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "745bfb78",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 493 ms, sys: 13.1 ms, total: 506 ms\n",
      "Wall time: 550 ms\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "_, titles = index.query_with_exclusions(tf.constant([random_user]), \n",
    "                                       tf.constant([items_to_exclude]))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "bdfb14a6",
   "metadata": {},
   "source": [
    "Now we can actually examine these recommendations to see if they make any sense! "
   ]
  },
  {
   "cell_type": "markdown",
   "id": "db96d721",
   "metadata": {},
   "source": [
    "**Historical purchases**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "f89702d2",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>user_no</th>\n",
       "      <th>item_no</th>\n",
       "      <th>gender_description</th>\n",
       "      <th>brand</th>\n",
       "      <th>product_group</th>\n",
       "      <th>first_interaction_month</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>1086</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>3644027680429875428</td>\n",
       "      <td>unisex</td>\n",
       "      <td>kuling</td>\n",
       "      <td>clothing sets</td>\n",
       "      <td>9</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>22976</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>7547780292943122816</td>\n",
       "      <td>girls</td>\n",
       "      <td>hype</td>\n",
       "      <td>bags</td>\n",
       "      <td>9</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>56639</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-488449987991724866</td>\n",
       "      <td>girls</td>\n",
       "      <td>molo</td>\n",
       "      <td>tops</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>78130</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>5072443198733158252</td>\n",
       "      <td>girls</td>\n",
       "      <td>hype</td>\n",
       "      <td>bags</td>\n",
       "      <td>9</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>127732</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-697862567626410763</td>\n",
       "      <td>girls</td>\n",
       "      <td>molo</td>\n",
       "      <td>bags</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>172484</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-932661838383765191</td>\n",
       "      <td>girls</td>\n",
       "      <td>kuling</td>\n",
       "      <td>coveralls</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>178817</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-50943422945763155</td>\n",
       "      <td>girls</td>\n",
       "      <td>hype</td>\n",
       "      <td>bags</td>\n",
       "      <td>9</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>203611</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-6961812644339906604</td>\n",
       "      <td>girls</td>\n",
       "      <td>emma och malena maternity</td>\n",
       "      <td>maternity tops</td>\n",
       "      <td>6</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>240005</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-7279020476320904084</td>\n",
       "      <td>girls</td>\n",
       "      <td>le toy van</td>\n",
       "      <td>role play</td>\n",
       "      <td>3</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>323177</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-7380651070414010617</td>\n",
       "      <td>unisex</td>\n",
       "      <td>kavat</td>\n",
       "      <td>sandals</td>\n",
       "      <td>3</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>325311</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>6023298784163360880</td>\n",
       "      <td>unisex</td>\n",
       "      <td>tretorn</td>\n",
       "      <td>boots</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>336312</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-2383862218621034937</td>\n",
       "      <td>unisex</td>\n",
       "      <td>jox</td>\n",
       "      <td>storage</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>385588</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-419121234283431680</td>\n",
       "      <td>unisex</td>\n",
       "      <td>kavat</td>\n",
       "      <td>sandals</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>407074</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>4498392745999421884</td>\n",
       "      <td>unisex</td>\n",
       "      <td>reima</td>\n",
       "      <td>coveralls</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>439040</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-64526162570614745</td>\n",
       "      <td>unisex</td>\n",
       "      <td>reima</td>\n",
       "      <td>coveralls</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>446751</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>6232313031656944598</td>\n",
       "      <td>unisex</td>\n",
       "      <td>teddykompaniet</td>\n",
       "      <td>soft toys</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>478904</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-2284907458364359995</td>\n",
       "      <td>girls</td>\n",
       "      <td>tenson</td>\n",
       "      <td>sandals</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>532637</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>8838709503091020009</td>\n",
       "      <td>unisex</td>\n",
       "      <td>viking</td>\n",
       "      <td>boots</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>589273</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-7765347603233238784</td>\n",
       "      <td>girls</td>\n",
       "      <td>viking</td>\n",
       "      <td>sandals</td>\n",
       "      <td>6</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>614409</th>\n",
       "      <td>-4227527154985019677</td>\n",
       "      <td>-7162062099089716900</td>\n",
       "      <td>girls</td>\n",
       "      <td>kuling</td>\n",
       "      <td>coveralls</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                     user_no               item_no gender_description  \\\n",
       "1086    -4227527154985019677   3644027680429875428             unisex   \n",
       "22976   -4227527154985019677   7547780292943122816              girls   \n",
       "56639   -4227527154985019677   -488449987991724866              girls   \n",
       "78130   -4227527154985019677   5072443198733158252              girls   \n",
       "127732  -4227527154985019677   -697862567626410763              girls   \n",
       "172484  -4227527154985019677   -932661838383765191              girls   \n",
       "178817  -4227527154985019677    -50943422945763155              girls   \n",
       "203611  -4227527154985019677  -6961812644339906604              girls   \n",
       "240005  -4227527154985019677  -7279020476320904084              girls   \n",
       "323177  -4227527154985019677  -7380651070414010617             unisex   \n",
       "325311  -4227527154985019677   6023298784163360880             unisex   \n",
       "336312  -4227527154985019677  -2383862218621034937             unisex   \n",
       "385588  -4227527154985019677   -419121234283431680             unisex   \n",
       "407074  -4227527154985019677   4498392745999421884             unisex   \n",
       "439040  -4227527154985019677    -64526162570614745             unisex   \n",
       "446751  -4227527154985019677   6232313031656944598             unisex   \n",
       "478904  -4227527154985019677  -2284907458364359995              girls   \n",
       "532637  -4227527154985019677   8838709503091020009             unisex   \n",
       "589273  -4227527154985019677  -7765347603233238784              girls   \n",
       "614409  -4227527154985019677  -7162062099089716900              girls   \n",
       "\n",
       "                            brand   product_group  first_interaction_month  \n",
       "1086                       kuling   clothing sets                        9  \n",
       "22976                        hype            bags                        9  \n",
       "56639                        molo            tops                       10  \n",
       "78130                        hype            bags                        9  \n",
       "127732                       molo            bags                       10  \n",
       "172484                     kuling       coveralls                        2  \n",
       "178817                       hype            bags                        9  \n",
       "203611  emma och malena maternity  maternity tops                        6  \n",
       "240005                 le toy van       role play                        3  \n",
       "323177                      kavat         sandals                        3  \n",
       "325311                    tretorn           boots                        2  \n",
       "336312                        jox         storage                       10  \n",
       "385588                      kavat         sandals                        2  \n",
       "407074                      reima       coveralls                        2  \n",
       "439040                      reima       coveralls                        2  \n",
       "446751             teddykompaniet       soft toys                       10  \n",
       "478904                     tenson         sandals                        2  \n",
       "532637                     viking           boots                       10  \n",
       "589273                     viking         sandals                        6  \n",
       "614409                     kuling       coveralls                        2  "
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_df_filtered.loc[train_df_filtered['user_no'] == random_user]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "96cf2b2b",
   "metadata": {},
   "source": [
    "**Recommendations**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "c2397d4e",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>item_no</th>\n",
       "      <th>colour</th>\n",
       "      <th>gender_description</th>\n",
       "      <th>brand</th>\n",
       "      <th>product_group</th>\n",
       "      <th>min_age</th>\n",
       "      <th>max_age</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>6819</th>\n",
       "      <td>6248379259425891960</td>\n",
       "      <td>yellow</td>\n",
       "      <td>unisex</td>\n",
       "      <td>mainio</td>\n",
       "      <td>bottoms</td>\n",
       "      <td>0.125</td>\n",
       "      <td>10.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>13381</th>\n",
       "      <td>40742030666431930</td>\n",
       "      <td>blue</td>\n",
       "      <td>unisex</td>\n",
       "      <td>kavat</td>\n",
       "      <td>boots</td>\n",
       "      <td>1.000</td>\n",
       "      <td>11.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>16889</th>\n",
       "      <td>7423339463580660099</td>\n",
       "      <td>white</td>\n",
       "      <td>unisex</td>\n",
       "      <td>mini rodini</td>\n",
       "      <td>fleeces and midlayers</td>\n",
       "      <td>0.375</td>\n",
       "      <td>11.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>25142</th>\n",
       "      <td>-2685666217734801200</td>\n",
       "      <td>pink</td>\n",
       "      <td>unisex</td>\n",
       "      <td>stoy</td>\n",
       "      <td>bicycles and other vehicles</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>35926</th>\n",
       "      <td>2649172241095163913</td>\n",
       "      <td>white</td>\n",
       "      <td>unisex</td>\n",
       "      <td>mini rodini</td>\n",
       "      <td>coats and jackets</td>\n",
       "      <td>0.875</td>\n",
       "      <td>11.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>36394</th>\n",
       "      <td>1011817660855101910</td>\n",
       "      <td>white</td>\n",
       "      <td>unisex</td>\n",
       "      <td>stoy</td>\n",
       "      <td>role play</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>47917</th>\n",
       "      <td>1052690307609796416</td>\n",
       "      <td>purple</td>\n",
       "      <td>unisex</td>\n",
       "      <td>kuling</td>\n",
       "      <td>boots</td>\n",
       "      <td>0.875</td>\n",
       "      <td>11.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>48272</th>\n",
       "      <td>7799576263764959168</td>\n",
       "      <td>blue</td>\n",
       "      <td>unisex</td>\n",
       "      <td>kuling</td>\n",
       "      <td>swimwear and coverups</td>\n",
       "      <td>0.625</td>\n",
       "      <td>6.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>51095</th>\n",
       "      <td>-6507239208000082480</td>\n",
       "      <td>black</td>\n",
       "      <td>unisex</td>\n",
       "      <td>kuling</td>\n",
       "      <td>boots</td>\n",
       "      <td>1.000</td>\n",
       "      <td>10.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>57234</th>\n",
       "      <td>-2121255597513549382</td>\n",
       "      <td>purple</td>\n",
       "      <td>unisex</td>\n",
       "      <td>kuling</td>\n",
       "      <td>coveralls</td>\n",
       "      <td>0.625</td>\n",
       "      <td>4.0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                    item_no  colour gender_description        brand  \\\n",
       "6819    6248379259425891960  yellow             unisex       mainio   \n",
       "13381     40742030666431930    blue             unisex        kavat   \n",
       "16889   7423339463580660099   white             unisex  mini rodini   \n",
       "25142  -2685666217734801200    pink             unisex         stoy   \n",
       "35926   2649172241095163913   white             unisex  mini rodini   \n",
       "36394   1011817660855101910   white             unisex         stoy   \n",
       "47917   1052690307609796416  purple             unisex       kuling   \n",
       "48272   7799576263764959168    blue             unisex       kuling   \n",
       "51095  -6507239208000082480   black             unisex       kuling   \n",
       "57234  -2121255597513549382  purple             unisex       kuling   \n",
       "\n",
       "                     product_group  min_age  max_age  \n",
       "6819                       bottoms    0.125     10.0  \n",
       "13381                        boots    1.000     11.0  \n",
       "16889        fleeces and midlayers    0.375     11.0  \n",
       "25142  bicycles and other vehicles      NaN      NaN  \n",
       "35926            coats and jackets    0.875     11.0  \n",
       "36394                    role play      NaN      NaN  \n",
       "47917                        boots    0.875     11.0  \n",
       "48272        swimwear and coverups    0.625      6.0  \n",
       "51095                        boots    1.000     10.0  \n",
       "57234                    coveralls    0.625      4.0  "
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "recommendations = [item.numpy().decode() for item in titles[0]]\n",
    "item_info_df.loc[item_info_df['item_no'].isin(recommendations)]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ec0f68b0",
   "metadata": {},
   "source": [
    "---\n",
    "---\n",
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a84e734a",
   "metadata": {},
   "source": [
    "## **Content-Based Filtering**\n",
    "\n",
    "Another way to approach recommendations is to base them solely on content metadata, rather than learning from patterns in interactions across the customer base as a whole. As such, we will likely not get any \"novel\" recommendations and instead many of the recommendations will be very similar to the user's purchase history. \n",
    "\n",
    "In order to take advantage of some TFRS \"machinery\", we can build user and item models as before. However, this time instead of *learning* embeddings for each individual user and each item, we will manually compute the representations of each user and each item. \n",
    "\n",
    "In this case an item embedding will just consist of the concatenated one-hot encodings of the brand, product group, and gender description, and a user embedding will be the average of all the item embeddings in their purchase history. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "id": "3dc746f3",
   "metadata": {},
   "outputs": [],
   "source": [
    "top_brands = get_vocab(train_df_filtered, 'brand', 100)\n",
    "top_groups = get_vocab(train_df_filtered, 'product_group', 50)\n",
    "COLS_TO_KEEP = ['gender_description', 'brand', 'product_group']\n",
    "\n",
    "def precompute_embeddings(df, agg_col):\n",
    "    df.loc[:, 'brand'] = df['brand'].apply(lambda x: x if x in top_brands else 'other')\n",
    "    df.loc[:, 'product_group'] = df['product_group'].apply(lambda x: x if x in top_groups else 'other')\n",
    "    df_one_hot = pd.get_dummies(df[COLS_TO_KEEP + [agg_col]], columns=COLS_TO_KEEP)\n",
    "    return df_one_hot.groupby(agg_col).agg('mean')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "id": "b1269a3a",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/thomasgaddy/Documents/Code/TFRecommenders/tfrs-retail-example/.venv/lib/python3.7/site-packages/pandas/core/indexing.py:964: SettingWithCopyWarning: \n",
      "A value is trying to be set on a copy of a slice from a DataFrame.\n",
      "Try using .loc[row_indexer,col_indexer] = value instead\n",
      "\n",
      "See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy\n",
      "  self.obj[item] = s\n"
     ]
    }
   ],
   "source": [
    "precomputed_user_embeddings = precompute_embeddings(train_df_filtered, agg_col='user_no')\n",
    "precomputed_item_embeddings = precompute_embeddings(item_info_df, agg_col='item_no')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "id": "47e24e9c",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>gender_description_boys</th>\n",
       "      <th>gender_description_girls</th>\n",
       "      <th>gender_description_unisex</th>\n",
       "      <th>brand_1+ in the family</th>\n",
       "      <th>brand_a happy brand</th>\n",
       "      <th>brand_a monday in copenhagen</th>\n",
       "      <th>brand_adidas</th>\n",
       "      <th>brand_barts</th>\n",
       "      <th>brand_beau loves</th>\n",
       "      <th>brand_billieblush</th>\n",
       "      <th>...</th>\n",
       "      <th>product_group_stroller parts and customisati</th>\n",
       "      <th>product_group_strollers</th>\n",
       "      <th>product_group_swimwear and coverups</th>\n",
       "      <th>product_group_tableware</th>\n",
       "      <th>product_group_textile</th>\n",
       "      <th>product_group_tops</th>\n",
       "      <th>product_group_trainers</th>\n",
       "      <th>product_group_underwear</th>\n",
       "      <th>product_group_vehicles</th>\n",
       "      <th>product_group_water toys</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>user_no</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>-1012876894217140776</th>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.277778</td>\n",
       "      <td>0.722222</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.166667</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.055556</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.055556</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.055556</td>\n",
       "      <td>0.166667</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>-1022934284196456562</th>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.166667</td>\n",
       "      <td>0.833333</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.111111</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.055556</td>\n",
       "      <td>0.111111</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>-1031375167955555195</th>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.736842</td>\n",
       "      <td>0.263158</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.210526</td>\n",
       "      <td>0.052632</td>\n",
       "      <td>0.105263</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>-1041412818309902183</th>\n",
       "      <td>0.200000</td>\n",
       "      <td>0.550000</td>\n",
       "      <td>0.250000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.600000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>-1044709512978776856</th>\n",
       "      <td>0.200000</td>\n",
       "      <td>0.050000</td>\n",
       "      <td>0.750000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.050000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.050000</td>\n",
       "      <td>0.050000</td>\n",
       "      <td>0.150000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>965297199758713016</th>\n",
       "      <td>0.166667</td>\n",
       "      <td>0.111111</td>\n",
       "      <td>0.722222</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.055556</td>\n",
       "      <td>0.055556</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.166667</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.055556</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>968073716034597193</th>\n",
       "      <td>0.111111</td>\n",
       "      <td>0.777778</td>\n",
       "      <td>0.111111</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.166667</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>976567085753614314</th>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.947368</td>\n",
       "      <td>0.052632</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.052632</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.052632</td>\n",
       "      <td>0.210526</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.157895</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>987479213534973896</th>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.277778</td>\n",
       "      <td>0.722222</td>\n",
       "      <td>0.166667</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.055556</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.277778</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>988004909956453914</th>\n",
       "      <td>0.050000</td>\n",
       "      <td>0.600000</td>\n",
       "      <td>0.350000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.100000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>2000 rows × 155 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "                      gender_description_boys  gender_description_girls  \\\n",
       "user_no                                                                   \n",
       "-1012876894217140776                 0.000000                  0.277778   \n",
       "-1022934284196456562                 0.000000                  0.166667   \n",
       "-1031375167955555195                 0.000000                  0.736842   \n",
       "-1041412818309902183                 0.200000                  0.550000   \n",
       "-1044709512978776856                 0.200000                  0.050000   \n",
       "...                                       ...                       ...   \n",
       "965297199758713016                   0.166667                  0.111111   \n",
       "968073716034597193                   0.111111                  0.777778   \n",
       "976567085753614314                   0.000000                  0.947368   \n",
       "987479213534973896                   0.000000                  0.277778   \n",
       "988004909956453914                   0.050000                  0.600000   \n",
       "\n",
       "                      gender_description_unisex  brand_1+ in the family  \\\n",
       "user_no                                                                   \n",
       "-1012876894217140776                   0.722222                0.000000   \n",
       "-1022934284196456562                   0.833333                0.000000   \n",
       "-1031375167955555195                   0.263158                0.000000   \n",
       "-1041412818309902183                   0.250000                0.000000   \n",
       "-1044709512978776856                   0.750000                0.000000   \n",
       "...                                         ...                     ...   \n",
       "965297199758713016                     0.722222                0.000000   \n",
       "968073716034597193                     0.111111                0.000000   \n",
       "976567085753614314                     0.052632                0.000000   \n",
       "987479213534973896                     0.722222                0.166667   \n",
       "988004909956453914                     0.350000                0.000000   \n",
       "\n",
       "                      brand_a happy brand  brand_a monday in copenhagen  \\\n",
       "user_no                                                                   \n",
       "-1012876894217140776             0.166667                      0.000000   \n",
       "-1022934284196456562             0.000000                      0.000000   \n",
       "-1031375167955555195             0.000000                      0.000000   \n",
       "-1041412818309902183             0.000000                      0.000000   \n",
       "-1044709512978776856             0.000000                      0.000000   \n",
       "...                                   ...                           ...   \n",
       "965297199758713016               0.000000                      0.000000   \n",
       "968073716034597193               0.000000                      0.000000   \n",
       "976567085753614314               0.000000                      0.052632   \n",
       "987479213534973896               0.000000                      0.000000   \n",
       "988004909956453914               0.000000                      0.000000   \n",
       "\n",
       "                      brand_adidas  brand_barts  brand_beau loves  \\\n",
       "user_no                                                             \n",
       "-1012876894217140776      0.055556          0.0               0.0   \n",
       "-1022934284196456562      0.000000          0.0               0.0   \n",
       "-1031375167955555195      0.000000          0.0               0.0   \n",
       "-1041412818309902183      0.000000          0.0               0.0   \n",
       "-1044709512978776856      0.050000          0.0               0.0   \n",
       "...                            ...          ...               ...   \n",
       "965297199758713016        0.000000          0.0               0.0   \n",
       "968073716034597193        0.000000          0.0               0.0   \n",
       "976567085753614314        0.000000          0.0               0.0   \n",
       "987479213534973896        0.000000          0.0               0.0   \n",
       "988004909956453914        0.000000          0.0               0.0   \n",
       "\n",
       "                      brand_billieblush  ...  \\\n",
       "user_no                                  ...   \n",
       "-1012876894217140776                0.0  ...   \n",
       "-1022934284196456562                0.0  ...   \n",
       "-1031375167955555195                0.0  ...   \n",
       "-1041412818309902183                0.0  ...   \n",
       "-1044709512978776856                0.0  ...   \n",
       "...                                 ...  ...   \n",
       "965297199758713016                  0.0  ...   \n",
       "968073716034597193                  0.0  ...   \n",
       "976567085753614314                  0.0  ...   \n",
       "987479213534973896                  0.0  ...   \n",
       "988004909956453914                  0.0  ...   \n",
       "\n",
       "                      product_group_stroller parts and customisati  \\\n",
       "user_no                                                              \n",
       "-1012876894217140776                                      0.000000   \n",
       "-1022934284196456562                                      0.000000   \n",
       "-1031375167955555195                                      0.000000   \n",
       "-1041412818309902183                                      0.000000   \n",
       "-1044709512978776856                                      0.000000   \n",
       "...                                                            ...   \n",
       "965297199758713016                                        0.055556   \n",
       "968073716034597193                                        0.000000   \n",
       "976567085753614314                                        0.000000   \n",
       "987479213534973896                                        0.000000   \n",
       "988004909956453914                                        0.000000   \n",
       "\n",
       "                      product_group_strollers  \\\n",
       "user_no                                         \n",
       "-1012876894217140776                 0.000000   \n",
       "-1022934284196456562                 0.000000   \n",
       "-1031375167955555195                 0.000000   \n",
       "-1041412818309902183                 0.000000   \n",
       "-1044709512978776856                 0.000000   \n",
       "...                                       ...   \n",
       "965297199758713016                   0.055556   \n",
       "968073716034597193                   0.000000   \n",
       "976567085753614314                   0.000000   \n",
       "987479213534973896                   0.000000   \n",
       "988004909956453914                   0.000000   \n",
       "\n",
       "                      product_group_swimwear and coverups  \\\n",
       "user_no                                                     \n",
       "-1012876894217140776                             0.055556   \n",
       "-1022934284196456562                             0.111111   \n",
       "-1031375167955555195                             0.000000   \n",
       "-1041412818309902183                             0.000000   \n",
       "-1044709512978776856                             0.000000   \n",
       "...                                                   ...   \n",
       "965297199758713016                               0.000000   \n",
       "968073716034597193                               0.000000   \n",
       "976567085753614314                               0.000000   \n",
       "987479213534973896                               0.000000   \n",
       "988004909956453914                               0.000000   \n",
       "\n",
       "                      product_group_tableware  product_group_textile  \\\n",
       "user_no                                                                \n",
       "-1012876894217140776                      0.0               0.000000   \n",
       "-1022934284196456562                      0.0               0.055556   \n",
       "-1031375167955555195                      0.0               0.000000   \n",
       "-1041412818309902183                      0.0               0.000000   \n",
       "-1044709512978776856                      0.0               0.000000   \n",
       "...                                       ...                    ...   \n",
       "965297199758713016                        0.0               0.000000   \n",
       "968073716034597193                        0.0               0.000000   \n",
       "976567085753614314                        0.0               0.052632   \n",
       "987479213534973896                        0.0               0.055556   \n",
       "988004909956453914                        0.0               0.000000   \n",
       "\n",
       "                      product_group_tops  product_group_trainers  \\\n",
       "user_no                                                            \n",
       "-1012876894217140776            0.000000                0.055556   \n",
       "-1022934284196456562            0.111111                0.000000   \n",
       "-1031375167955555195            0.210526                0.052632   \n",
       "-1041412818309902183            0.600000                0.000000   \n",
       "-1044709512978776856            0.050000                0.050000   \n",
       "...                                  ...                     ...   \n",
       "965297199758713016              0.166667                0.000000   \n",
       "968073716034597193              0.166667                0.000000   \n",
       "976567085753614314              0.210526                0.000000   \n",
       "987479213534973896              0.000000                0.000000   \n",
       "988004909956453914              0.000000                0.100000   \n",
       "\n",
       "                      product_group_underwear  product_group_vehicles  \\\n",
       "user_no                                                                 \n",
       "-1012876894217140776                 0.166667                     0.0   \n",
       "-1022934284196456562                 0.000000                     0.0   \n",
       "-1031375167955555195                 0.105263                     0.0   \n",
       "-1041412818309902183                 0.000000                     0.0   \n",
       "-1044709512978776856                 0.150000                     0.0   \n",
       "...                                       ...                     ...   \n",
       "965297199758713016                   0.055556                     0.0   \n",
       "968073716034597193                   0.000000                     0.0   \n",
       "976567085753614314                   0.157895                     0.0   \n",
       "987479213534973896                   0.277778                     0.0   \n",
       "988004909956453914                   0.000000                     0.0   \n",
       "\n",
       "                      product_group_water toys  \n",
       "user_no                                         \n",
       "-1012876894217140776                       0.0  \n",
       "-1022934284196456562                       0.0  \n",
       "-1031375167955555195                       0.0  \n",
       "-1041412818309902183                       0.0  \n",
       "-1044709512978776856                       0.0  \n",
       "...                                        ...  \n",
       "965297199758713016                         0.0  \n",
       "968073716034597193                         0.0  \n",
       "976567085753614314                         0.0  \n",
       "987479213534973896                         0.0  \n",
       "988004909956453914                         0.0  \n",
       "\n",
       "[2000 rows x 155 columns]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "display(precomputed_user_embeddings)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "id": "7782e626",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>gender_description_boys</th>\n",
       "      <th>gender_description_girls</th>\n",
       "      <th>gender_description_unisex</th>\n",
       "      <th>brand_1+ in the family</th>\n",
       "      <th>brand_a happy brand</th>\n",
       "      <th>brand_a monday in copenhagen</th>\n",
       "      <th>brand_adidas</th>\n",
       "      <th>brand_barts</th>\n",
       "      <th>brand_beau loves</th>\n",
       "      <th>brand_billieblush</th>\n",
       "      <th>...</th>\n",
       "      <th>product_group_stroller parts and customisati</th>\n",
       "      <th>product_group_strollers</th>\n",
       "      <th>product_group_swimwear and coverups</th>\n",
       "      <th>product_group_tableware</th>\n",
       "      <th>product_group_textile</th>\n",
       "      <th>product_group_tops</th>\n",
       "      <th>product_group_trainers</th>\n",
       "      <th>product_group_underwear</th>\n",
       "      <th>product_group_vehicles</th>\n",
       "      <th>product_group_water toys</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>item_no</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>-10001501373726678</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>-1000182030290830232</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>-1000183384954605528</th>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>-1000321715684049686</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>-1000570342615087077</th>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>999030474988862413</th>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>999032067904529387</th>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>999084409713144028</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>999328979874402204</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>999657375817052613</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>61704 rows × 155 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "                      gender_description_boys  gender_description_girls  \\\n",
       "item_no                                                                   \n",
       "-10001501373726678                          0                         0   \n",
       "-1000182030290830232                        0                         0   \n",
       "-1000183384954605528                        0                         1   \n",
       "-1000321715684049686                        0                         0   \n",
       "-1000570342615087077                        1                         0   \n",
       "...                                       ...                       ...   \n",
       "999030474988862413                          1                         0   \n",
       "999032067904529387                          0                         1   \n",
       "999084409713144028                          0                         0   \n",
       "999328979874402204                          0                         0   \n",
       "999657375817052613                          0                         0   \n",
       "\n",
       "                      gender_description_unisex  brand_1+ in the family  \\\n",
       "item_no                                                                   \n",
       "-10001501373726678                            1                       0   \n",
       "-1000182030290830232                          1                       0   \n",
       "-1000183384954605528                          0                       0   \n",
       "-1000321715684049686                          1                       0   \n",
       "-1000570342615087077                          0                       0   \n",
       "...                                         ...                     ...   \n",
       "999030474988862413                            0                       0   \n",
       "999032067904529387                            0                       0   \n",
       "999084409713144028                            1                       0   \n",
       "999328979874402204                            1                       0   \n",
       "999657375817052613                            1                       0   \n",
       "\n",
       "                      brand_a happy brand  brand_a monday in copenhagen  \\\n",
       "item_no                                                                   \n",
       "-10001501373726678                      0                             0   \n",
       "-1000182030290830232                    0                             0   \n",
       "-1000183384954605528                    0                             0   \n",
       "-1000321715684049686                    0                             0   \n",
       "-1000570342615087077                    0                             0   \n",
       "...                                   ...                           ...   \n",
       "999030474988862413                      0                             0   \n",
       "999032067904529387                      0                             0   \n",
       "999084409713144028                      0                             0   \n",
       "999328979874402204                      0                             0   \n",
       "999657375817052613                      0                             0   \n",
       "\n",
       "                      brand_adidas  brand_barts  brand_beau loves  \\\n",
       "item_no                                                             \n",
       "-10001501373726678               0            0                 0   \n",
       "-1000182030290830232             0            0                 0   \n",
       "-1000183384954605528             0            0                 0   \n",
       "-1000321715684049686             0            0                 0   \n",
       "-1000570342615087077             0            0                 0   \n",
       "...                            ...          ...               ...   \n",
       "999030474988862413               0            0                 0   \n",
       "999032067904529387               0            0                 0   \n",
       "999084409713144028               0            0                 0   \n",
       "999328979874402204               0            0                 0   \n",
       "999657375817052613               0            0                 0   \n",
       "\n",
       "                      brand_billieblush  ...  \\\n",
       "item_no                                  ...   \n",
       "-10001501373726678                    0  ...   \n",
       "-1000182030290830232                  0  ...   \n",
       "-1000183384954605528                  0  ...   \n",
       "-1000321715684049686                  0  ...   \n",
       "-1000570342615087077                  0  ...   \n",
       "...                                 ...  ...   \n",
       "999030474988862413                    0  ...   \n",
       "999032067904529387                    0  ...   \n",
       "999084409713144028                    0  ...   \n",
       "999328979874402204                    0  ...   \n",
       "999657375817052613                    0  ...   \n",
       "\n",
       "                      product_group_stroller parts and customisati  \\\n",
       "item_no                                                              \n",
       "-10001501373726678                                               0   \n",
       "-1000182030290830232                                             0   \n",
       "-1000183384954605528                                             0   \n",
       "-1000321715684049686                                             0   \n",
       "-1000570342615087077                                             0   \n",
       "...                                                            ...   \n",
       "999030474988862413                                               0   \n",
       "999032067904529387                                               0   \n",
       "999084409713144028                                               0   \n",
       "999328979874402204                                               0   \n",
       "999657375817052613                                               0   \n",
       "\n",
       "                      product_group_strollers  \\\n",
       "item_no                                         \n",
       "-10001501373726678                          0   \n",
       "-1000182030290830232                        0   \n",
       "-1000183384954605528                        0   \n",
       "-1000321715684049686                        0   \n",
       "-1000570342615087077                        0   \n",
       "...                                       ...   \n",
       "999030474988862413                          0   \n",
       "999032067904529387                          0   \n",
       "999084409713144028                          0   \n",
       "999328979874402204                          0   \n",
       "999657375817052613                          0   \n",
       "\n",
       "                      product_group_swimwear and coverups  \\\n",
       "item_no                                                     \n",
       "-10001501373726678                                      0   \n",
       "-1000182030290830232                                    0   \n",
       "-1000183384954605528                                    0   \n",
       "-1000321715684049686                                    0   \n",
       "-1000570342615087077                                    0   \n",
       "...                                                   ...   \n",
       "999030474988862413                                      0   \n",
       "999032067904529387                                      0   \n",
       "999084409713144028                                      0   \n",
       "999328979874402204                                      0   \n",
       "999657375817052613                                      0   \n",
       "\n",
       "                      product_group_tableware  product_group_textile  \\\n",
       "item_no                                                                \n",
       "-10001501373726678                          0                      0   \n",
       "-1000182030290830232                        0                      0   \n",
       "-1000183384954605528                        0                      0   \n",
       "-1000321715684049686                        0                      0   \n",
       "-1000570342615087077                        0                      0   \n",
       "...                                       ...                    ...   \n",
       "999030474988862413                          0                      0   \n",
       "999032067904529387                          0                      0   \n",
       "999084409713144028                          0                      0   \n",
       "999328979874402204                          0                      0   \n",
       "999657375817052613                          0                      0   \n",
       "\n",
       "                      product_group_tops  product_group_trainers  \\\n",
       "item_no                                                            \n",
       "-10001501373726678                     0                       0   \n",
       "-1000182030290830232                   0                       1   \n",
       "-1000183384954605528                   0                       0   \n",
       "-1000321715684049686                   0                       0   \n",
       "-1000570342615087077                   0                       0   \n",
       "...                                  ...                     ...   \n",
       "999030474988862413                     1                       0   \n",
       "999032067904529387                     0                       0   \n",
       "999084409713144028                     0                       0   \n",
       "999328979874402204                     0                       0   \n",
       "999657375817052613                     0                       0   \n",
       "\n",
       "                      product_group_underwear  product_group_vehicles  \\\n",
       "item_no                                                                 \n",
       "-10001501373726678                          0                       0   \n",
       "-1000182030290830232                        0                       0   \n",
       "-1000183384954605528                        0                       0   \n",
       "-1000321715684049686                        0                       0   \n",
       "-1000570342615087077                        0                       0   \n",
       "...                                       ...                     ...   \n",
       "999030474988862413                          0                       0   \n",
       "999032067904529387                          0                       0   \n",
       "999084409713144028                          0                       0   \n",
       "999328979874402204                          0                       0   \n",
       "999657375817052613                          0                       0   \n",
       "\n",
       "                      product_group_water toys  \n",
       "item_no                                         \n",
       "-10001501373726678                           0  \n",
       "-1000182030290830232                         0  \n",
       "-1000183384954605528                         0  \n",
       "-1000321715684049686                         0  \n",
       "-1000570342615087077                         0  \n",
       "...                                        ...  \n",
       "999030474988862413                           0  \n",
       "999032067904529387                           0  \n",
       "999084409713144028                           0  \n",
       "999328979874402204                           0  \n",
       "999657375817052613                           0  \n",
       "\n",
       "[61704 rows x 155 columns]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "display(precomputed_item_embeddings)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "id": "94c1d6ab",
   "metadata": {},
   "outputs": [],
   "source": [
    "def create_precomputed_embedding_model(precomputed_embeddings):\n",
    "    num_columns = len(precomputed_embeddings.columns)\n",
    "    embedding_matrix = np.concatenate((np.zeros((1, num_columns)),\n",
    "                                      precomputed_embeddings.values))\n",
    "    embedding_layer = tf.keras.layers.Embedding(*embedding_matrix.shape,\n",
    "                                                embeddings_initializer=tf.keras.initializers.Constant(\n",
    "                                                    embedding_matrix),\n",
    "                                                trainable=False)\n",
    "    model = tf.keras.Sequential([\n",
    "        tf.keras.layers.StringLookup(\n",
    "            vocabulary=precomputed_embeddings.index,\n",
    "            num_oov_indices=1\n",
    "        ),\n",
    "        embedding_layer\n",
    "    ])\n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "id": "ae414a77",
   "metadata": {},
   "outputs": [],
   "source": [
    "user_model = create_precomputed_embedding_model(precomputed_user_embeddings)\n",
    "item_model = create_precomputed_embedding_model(precomputed_item_embeddings)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a3e7f45d",
   "metadata": {},
   "source": [
    "The following cell demonstrates that all the item model is doing is looking up the one-hot encoding of brand, product group, and gender description for a user. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "id": "c0bac964",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<tf.Tensor: shape=(1, 155), dtype=float32, numpy=\n",
       "array([[0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n",
       "        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n",
       "        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n",
       "        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n",
       "        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n",
       "        0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0.,\n",
       "        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n",
       "        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n",
       "        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0.,\n",
       "        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]], dtype=float32)>"
      ]
     },
     "execution_count": 31,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "item_model(tf.constant(['-1000183384954605528']))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ecf696ed",
   "metadata": {},
   "source": [
    "Even though we didn't train a `tfrs.Model`, we can still create an index exactly as we did above! "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "id": "46ca4d6d",
   "metadata": {},
   "outputs": [],
   "source": [
    "items_dataset = tf.data.Dataset.from_tensor_slices(item_info_df['item_no'])\n",
    "# Create a model that takes in raw query features, and\n",
    "index = tfrs.layers.factorized_top_k.BruteForce(user_model)\n",
    "# recommends items out of the entire items dataset.\n",
    "_ = index.index_from_dataset(\n",
    "  tf.data.Dataset.zip((items_dataset.batch(100), items_dataset.batch(100).map(item_model)))\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "id": "519f769f",
   "metadata": {},
   "outputs": [],
   "source": [
    "random_user = np.random.choice(train_df_filtered['user_no'].unique())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "id": "aeca94a8",
   "metadata": {},
   "outputs": [],
   "source": [
    "items_to_exclude = train_df_filtered.loc[train_df_filtered['user_no'] == random_user]['item_no'].unique()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "id": "ab89ebf5",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 189 ms, sys: 7.29 ms, total: 196 ms\n",
      "Wall time: 211 ms\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "_, titles = index.query_with_exclusions(tf.constant([random_user]), \n",
    "                                       tf.constant([items_to_exclude]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "id": "0cd43c89",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>user_no</th>\n",
       "      <th>item_no</th>\n",
       "      <th>gender_description</th>\n",
       "      <th>brand</th>\n",
       "      <th>product_group</th>\n",
       "      <th>first_interaction_month</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>44764</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>-7062902217557939573</td>\n",
       "      <td>unisex</td>\n",
       "      <td>reima</td>\n",
       "      <td>headwear</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>61710</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>2164894343335326037</td>\n",
       "      <td>unisex</td>\n",
       "      <td>tartine et chocolat</td>\n",
       "      <td>shoes</td>\n",
       "      <td>9</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>98899</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>3933678542901544177</td>\n",
       "      <td>boys</td>\n",
       "      <td>molo</td>\n",
       "      <td>jumpers and knitwear</td>\n",
       "      <td>9</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>103112</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>4490900546980254270</td>\n",
       "      <td>boys</td>\n",
       "      <td>jacadi</td>\n",
       "      <td>jumpers and knitwear</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>178498</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>1193766607815201020</td>\n",
       "      <td>boys</td>\n",
       "      <td>other</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>9</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>181962</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>-4251817304791664609</td>\n",
       "      <td>unisex</td>\n",
       "      <td>fub</td>\n",
       "      <td>headwear</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>187693</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>-8038060062925076296</td>\n",
       "      <td>boys</td>\n",
       "      <td>other</td>\n",
       "      <td>trainers</td>\n",
       "      <td>3</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>195508</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>-5170604594377873809</td>\n",
       "      <td>girls</td>\n",
       "      <td>billieblush</td>\n",
       "      <td>tops</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>201915</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>-809248011627862386</td>\n",
       "      <td>girls</td>\n",
       "      <td>jacadi</td>\n",
       "      <td>dresses</td>\n",
       "      <td>9</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>214200</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>1305137013358548518</td>\n",
       "      <td>unisex</td>\n",
       "      <td>other</td>\n",
       "      <td>trainers</td>\n",
       "      <td>3</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>226885</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>6032649609097040334</td>\n",
       "      <td>unisex</td>\n",
       "      <td>stoy</td>\n",
       "      <td>vehicles</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>290494</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>-995432985202626179</td>\n",
       "      <td>unisex</td>\n",
       "      <td>fub</td>\n",
       "      <td>headwear</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>365281</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>-6990630456318014787</td>\n",
       "      <td>unisex</td>\n",
       "      <td>other</td>\n",
       "      <td>vehicles</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>421406</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>4329331830954320591</td>\n",
       "      <td>girls</td>\n",
       "      <td>jacadi</td>\n",
       "      <td>dresses</td>\n",
       "      <td>9</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>483070</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>5841045138243478923</td>\n",
       "      <td>boys</td>\n",
       "      <td>ralph lauren</td>\n",
       "      <td>tops</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>483112</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>-318792259142847741</td>\n",
       "      <td>unisex</td>\n",
       "      <td>marmar copenhagen</td>\n",
       "      <td>boots</td>\n",
       "      <td>9</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>513216</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>-7258898587632840201</td>\n",
       "      <td>boys</td>\n",
       "      <td>jacadi</td>\n",
       "      <td>coveralls</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>533771</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>-92825680858531790</td>\n",
       "      <td>unisex</td>\n",
       "      <td>kuling</td>\n",
       "      <td>fleeces and midlayers</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>663224</th>\n",
       "      <td>-582134469683447242</td>\n",
       "      <td>-2355774442937251683</td>\n",
       "      <td>unisex</td>\n",
       "      <td>reima</td>\n",
       "      <td>headwear</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                    user_no               item_no gender_description  \\\n",
       "44764   -582134469683447242  -7062902217557939573             unisex   \n",
       "61710   -582134469683447242   2164894343335326037             unisex   \n",
       "98899   -582134469683447242   3933678542901544177               boys   \n",
       "103112  -582134469683447242   4490900546980254270               boys   \n",
       "178498  -582134469683447242   1193766607815201020               boys   \n",
       "181962  -582134469683447242  -4251817304791664609             unisex   \n",
       "187693  -582134469683447242  -8038060062925076296               boys   \n",
       "195508  -582134469683447242  -5170604594377873809              girls   \n",
       "201915  -582134469683447242   -809248011627862386              girls   \n",
       "214200  -582134469683447242   1305137013358548518             unisex   \n",
       "226885  -582134469683447242   6032649609097040334             unisex   \n",
       "290494  -582134469683447242   -995432985202626179             unisex   \n",
       "365281  -582134469683447242  -6990630456318014787             unisex   \n",
       "421406  -582134469683447242   4329331830954320591              girls   \n",
       "483070  -582134469683447242   5841045138243478923               boys   \n",
       "483112  -582134469683447242   -318792259142847741             unisex   \n",
       "513216  -582134469683447242  -7258898587632840201               boys   \n",
       "533771  -582134469683447242    -92825680858531790             unisex   \n",
       "663224  -582134469683447242  -2355774442937251683             unisex   \n",
       "\n",
       "                      brand          product_group  first_interaction_month  \n",
       "44764                 reima               headwear                       11  \n",
       "61710   tartine et chocolat                  shoes                        9  \n",
       "98899                  molo   jumpers and knitwear                        9  \n",
       "103112               jacadi   jumpers and knitwear                       11  \n",
       "178498                other            all in ones                        9  \n",
       "181962                  fub               headwear                       11  \n",
       "187693                other               trainers                        3  \n",
       "195508          billieblush                   tops                       10  \n",
       "201915               jacadi                dresses                        9  \n",
       "214200                other               trainers                        3  \n",
       "226885                 stoy               vehicles                       10  \n",
       "290494                  fub               headwear                       11  \n",
       "365281                other               vehicles                       10  \n",
       "421406               jacadi                dresses                        9  \n",
       "483070         ralph lauren                   tops                       10  \n",
       "483112    marmar copenhagen                  boots                        9  \n",
       "513216               jacadi              coveralls                       11  \n",
       "533771               kuling  fleeces and midlayers                       11  \n",
       "663224                reima               headwear                       11  "
      ]
     },
     "execution_count": 36,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_df_filtered.loc[train_df_filtered['user_no'] == random_user]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "id": "28166995",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>item_no</th>\n",
       "      <th>colour</th>\n",
       "      <th>gender_description</th>\n",
       "      <th>brand</th>\n",
       "      <th>product_group</th>\n",
       "      <th>min_age</th>\n",
       "      <th>max_age</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>15</th>\n",
       "      <td>457467103957514638</td>\n",
       "      <td>beige</td>\n",
       "      <td>unisex</td>\n",
       "      <td>other</td>\n",
       "      <td>headwear</td>\n",
       "      <td>0.125</td>\n",
       "      <td>0.125</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>19</th>\n",
       "      <td>5751912426656680356</td>\n",
       "      <td>white</td>\n",
       "      <td>unisex</td>\n",
       "      <td>other</td>\n",
       "      <td>headwear</td>\n",
       "      <td>0.125</td>\n",
       "      <td>2.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>209</th>\n",
       "      <td>3958544075576929766</td>\n",
       "      <td>beige</td>\n",
       "      <td>unisex</td>\n",
       "      <td>other</td>\n",
       "      <td>headwear</td>\n",
       "      <td>0.625</td>\n",
       "      <td>8.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>252</th>\n",
       "      <td>5765193949212180695</td>\n",
       "      <td>yellow</td>\n",
       "      <td>unisex</td>\n",
       "      <td>other</td>\n",
       "      <td>headwear</td>\n",
       "      <td>0.375</td>\n",
       "      <td>2.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>495</th>\n",
       "      <td>7912340583694710060</td>\n",
       "      <td>grey</td>\n",
       "      <td>unisex</td>\n",
       "      <td>other</td>\n",
       "      <td>headwear</td>\n",
       "      <td>0.125</td>\n",
       "      <td>2.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>670</th>\n",
       "      <td>-8259248676814643686</td>\n",
       "      <td>black</td>\n",
       "      <td>unisex</td>\n",
       "      <td>other</td>\n",
       "      <td>headwear</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>942</th>\n",
       "      <td>2589311358657624625</td>\n",
       "      <td>black</td>\n",
       "      <td>unisex</td>\n",
       "      <td>other</td>\n",
       "      <td>headwear</td>\n",
       "      <td>0.125</td>\n",
       "      <td>4.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1022</th>\n",
       "      <td>5146575204438291293</td>\n",
       "      <td>grey</td>\n",
       "      <td>unisex</td>\n",
       "      <td>other</td>\n",
       "      <td>headwear</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1092</th>\n",
       "      <td>7732139752943756076</td>\n",
       "      <td>white</td>\n",
       "      <td>unisex</td>\n",
       "      <td>other</td>\n",
       "      <td>headwear</td>\n",
       "      <td>0.125</td>\n",
       "      <td>2.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1406</th>\n",
       "      <td>1088769495416475325</td>\n",
       "      <td>blue</td>\n",
       "      <td>unisex</td>\n",
       "      <td>other</td>\n",
       "      <td>headwear</td>\n",
       "      <td>0.125</td>\n",
       "      <td>0.875</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                   item_no  colour gender_description  brand product_group  \\\n",
       "15      457467103957514638   beige             unisex  other      headwear   \n",
       "19     5751912426656680356   white             unisex  other      headwear   \n",
       "209    3958544075576929766   beige             unisex  other      headwear   \n",
       "252    5765193949212180695  yellow             unisex  other      headwear   \n",
       "495    7912340583694710060    grey             unisex  other      headwear   \n",
       "670   -8259248676814643686   black             unisex  other      headwear   \n",
       "942    2589311358657624625   black             unisex  other      headwear   \n",
       "1022   5146575204438291293    grey             unisex  other      headwear   \n",
       "1092   7732139752943756076   white             unisex  other      headwear   \n",
       "1406   1088769495416475325    blue             unisex  other      headwear   \n",
       "\n",
       "      min_age  max_age  \n",
       "15      0.125    0.125  \n",
       "19      0.125    2.000  \n",
       "209     0.625    8.000  \n",
       "252     0.375    2.000  \n",
       "495     0.125    2.000  \n",
       "670       NaN      NaN  \n",
       "942     0.125    4.000  \n",
       "1022      NaN      NaN  \n",
       "1092    0.125    2.000  \n",
       "1406    0.125    0.875  "
      ]
     },
     "execution_count": 37,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "recommendations = [item.numpy().decode() for item in titles[0]]\n",
    "item_info_df.loc[item_info_df['item_no'].isin(recommendations)]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "eab7717a",
   "metadata": {},
   "source": [
    "**Usually the recommendations have little or no diversity**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "id": "111af655",
   "metadata": {},
   "outputs": [],
   "source": [
    "test_users_dataset = tf.data.Dataset.from_tensor_slices(test_df_filtered['user_no'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "id": "f4e00863",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 2.57 s, sys: 448 ms, total: 3.02 s\n",
      "Wall time: 1.28 s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "_, retrieved_items = index(tf.constant(test_df_filtered['user_no']), k=100)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "83cb5208",
   "metadata": {},
   "source": [
    "~1s to produce 100 recommendations for ~2000 users. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "id": "1d09abf8",
   "metadata": {},
   "outputs": [],
   "source": [
    "ids_match = tf.cast(tf.math.equal(true_candidates, retrieved_items), tf.float32)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "id": "6a0164d9",
   "metadata": {},
   "outputs": [],
   "source": [
    "metrics = [tf.keras.metrics.Mean() for k in ks]\n",
    "for k, metric in zip(ks, metrics):\n",
    "    # By slicing until :k we assume scores are sorted.\n",
    "    # Clip to only count multiple matches once.\n",
    "    match_found = tf.clip_by_value(\n",
    "        tf.reduce_sum(ids_match[:, :k], axis=1, keepdims=True),\n",
    "        0.0, 1.0\n",
    "    )\n",
    "    metric.update_state(match_found)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "id": "f938dd84",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Top 1 categorical accuracy: 0.00044\n",
      "Top 5 categorical accuracy: 0.00262\n",
      "Top 10 categorical accuracy: 0.00698\n",
      "Top 50 categorical accuracy: 0.03752\n",
      "Top 100 categorical accuracy: 0.06937\n"
     ]
    }
   ],
   "source": [
    "for k, metric in zip(ks, metrics):\n",
    "    print(f'Top {k} categorical accuracy: {metric.result().numpy():.5f}')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5f073908",
   "metadata": {},
   "source": [
    "This very simple model actually does much better than either the baseline or our simple collaborative filtering model. \n",
    "\n",
    "What if there was a way to get the best of both worlds?"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "792b6600",
   "metadata": {},
   "source": [
    "---\n",
    "\n",
    "# **Using Additional Features to Create a Hybrid Model**\n",
    "\n",
    "From the [using rich features tutorial](https://www.tensorflow.org/recommenders/examples/featurization): \n",
    "\n",
    "> One of the great advantages of using a deep learning framework to build recommender models is the freedom to build rich, flexible feature representations.\n",
    "\n",
    "We can use additional features to build our representations. In this case, we will use the additional metadata features we used in our content based filtering \"model\" (brand, product group, gender description) to create richer representations of each item. Intuitively, the collaborative filtering model created item representations solely based on purchasing behavior (\"many customers bought items A, B, and C, so their representations should be 'similar'. If a customer has bought items A and B, we should recommend C\"). On the other hand, the content based filtering model created item representations solely based on item metadata (\"items X, Y, and Z all have the same brand, product description, and gender description, so their representations should be 'similar'. If a customer has bought items X and Y, we should recommend Z\"). \n",
    "\n",
    "With this next model, we will try to combine these ideas so that item representations capture **both** purchasing behavior and item characteristics based on metadata features. \n",
    "\n",
    "We will only add additional features to the item model, but there is nothing stopping us from adding to the user model as well. If, for example, we had demographic features they could be worth adding to the user model. We could also add **context** features to the user/query model, such as time, device, etc. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "id": "a0b0e7d0",
   "metadata": {},
   "outputs": [],
   "source": [
    "class ItemModel(tf.keras.Model):\n",
    "    def __init__(self, \n",
    "                 train_df):\n",
    "        super().__init__()\n",
    "        # Embed the item id into 16 dimensions\n",
    "        self.item_embedding = create_embedding_model(train_df, \"item_no\", embedding_dim=16)\n",
    "\n",
    "        # There are three gender description values in the data, just one hot encode them\n",
    "        gender_description = get_vocab(train_df, 'gender_description')\n",
    "        self.gender_description_lookup = tf.keras.layers.StringLookup(vocabulary=gender_description, \n",
    "                                                                      output_mode='one_hot',\n",
    "                                                                      num_oov_indices=0)\n",
    "        self.brand_embedding = create_embedding_model(train_df, \n",
    "                                                     \"brand\", \n",
    "                                                     embedding_dim=8)\n",
    "        self.product_group_embedding = create_embedding_model(train_df, \n",
    "                                                              \"product_group\", \n",
    "                                                              embedding_dim=5)\n",
    "        \n",
    "    def call(self, inputs):\n",
    "        \"\"\"\n",
    "        Item representation is the concatenation of ID embedding, gender description one-hot \n",
    "        encoding, brand embedding, and product group embedding\n",
    "        \"\"\"\n",
    "        return tf.concat([\n",
    "             self.item_embedding(inputs['item_no']),\n",
    "             self.gender_description_lookup(inputs['gender_description']),\n",
    "             self.brand_embedding(inputs['brand']),\n",
    "             self.product_group_embedding(inputs['product_group'])\n",
    "        ], axis=1)\n",
    "    \n",
    "class TFRSContextModel(tfrs.models.Model):\n",
    "    def __init__(self, \n",
    "                 user_model,\n",
    "                 item_model, \n",
    "                 items_w_context):\n",
    "        super().__init__()\n",
    "        self.user_model = user_model\n",
    "        self.item_model = item_model\n",
    "        self.task = tfrs.tasks.Retrieval(\n",
    "            metrics = tfrs.metrics.FactorizedTopK(\n",
    "                candidates=items_w_context.batch(128).map(self.item_model)\n",
    "            )\n",
    "        )\n",
    "        \n",
    "    def compute_loss(self, inputs, training=False):\n",
    "        query_embeddings = self.user_model(inputs['user_no'])\n",
    "        candidate_embeddings = self.item_model({\n",
    "            'item_no': inputs['item_no'],\n",
    "            'gender_description': inputs['gender_description'],\n",
    "            'brand': inputs['brand'],\n",
    "            'product_group': inputs['product_group']\n",
    "        })\n",
    "        \n",
    "        return self.task(query_embeddings, candidate_embeddings, compute_metrics=not training)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "id": "07145a97",
   "metadata": {},
   "outputs": [],
   "source": [
    "items_df = item_info_df.loc[item_info_df['item_no'].isin(items)][\n",
    "    ['item_no', 'gender_description', 'brand', 'product_group']]\n",
    "\n",
    "items_dataset_w_context = tf.data.Dataset.from_tensor_slices(dict(items_df))\n",
    "user_model = create_embedding_model(train_df_filtered, \"user_no\")\n",
    "item_model = ItemModel(train_df_filtered)\n",
    "\n",
    "model = TFRSContextModel(user_model, item_model, items_dataset_w_context)\n",
    "model.compile(optimizer=tf.keras.optimizers.Adagrad(0.1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "id": "99c4c980",
   "metadata": {},
   "outputs": [],
   "source": [
    "train_ds = train_dataset.shuffle(1_000).batch(4096)\n",
    "test_ds = test_dataset.batch(4096)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "id": "04643ab9",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/3\n",
      "10/10 [==============================] - 8s 779ms/step - factorized_top_k/top_1_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_5_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_10_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_50_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_100_categorical_accuracy: 0.0000e+00 - loss: 27335.1411 - regularization_loss: 0.0000e+00 - total_loss: 27335.1411 - val_factorized_top_k/top_1_categorical_accuracy: 0.0201 - val_factorized_top_k/top_5_categorical_accuracy: 0.0249 - val_factorized_top_k/top_10_categorical_accuracy: 0.0310 - val_factorized_top_k/top_50_categorical_accuracy: 0.0851 - val_factorized_top_k/top_100_categorical_accuracy: 0.1239 - val_loss: 16602.7422 - val_regularization_loss: 0.0000e+00 - val_total_loss: 16602.7422\n",
      "Epoch 2/3\n",
      "10/10 [==============================] - 6s 589ms/step - factorized_top_k/top_1_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_5_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_10_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_50_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_100_categorical_accuracy: 0.0000e+00 - loss: 25110.3476 - regularization_loss: 0.0000e+00 - total_loss: 25110.3476 - val_factorized_top_k/top_1_categorical_accuracy: 0.0070 - val_factorized_top_k/top_5_categorical_accuracy: 0.0157 - val_factorized_top_k/top_10_categorical_accuracy: 0.0240 - val_factorized_top_k/top_50_categorical_accuracy: 0.0794 - val_factorized_top_k/top_100_categorical_accuracy: 0.1217 - val_loss: 15821.6846 - val_regularization_loss: 0.0000e+00 - val_total_loss: 15821.6846\n",
      "Epoch 3/3\n",
      "10/10 [==============================] - 6s 662ms/step - factorized_top_k/top_1_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_5_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_10_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_50_categorical_accuracy: 0.0000e+00 - factorized_top_k/top_100_categorical_accuracy: 0.0000e+00 - loss: 23554.3557 - regularization_loss: 0.0000e+00 - total_loss: 23554.3557 - val_factorized_top_k/top_1_categorical_accuracy: 0.0022 - val_factorized_top_k/top_5_categorical_accuracy: 0.0118 - val_factorized_top_k/top_10_categorical_accuracy: 0.0214 - val_factorized_top_k/top_50_categorical_accuracy: 0.0764 - val_factorized_top_k/top_100_categorical_accuracy: 0.1252 - val_loss: 15575.5381 - val_regularization_loss: 0.0000e+00 - val_total_loss: 15575.5381\n"
     ]
    }
   ],
   "source": [
    "history = model.fit(train_ds, epochs=3, validation_data=test_ds)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "id": "4d35c3a0",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "10/10 [==============================] - 55s 5s/step - factorized_top_k/top_1_categorical_accuracy: 0.0377 - factorized_top_k/top_5_categorical_accuracy: 0.1208 - factorized_top_k/top_10_categorical_accuracy: 0.1611 - factorized_top_k/top_50_categorical_accuracy: 0.2748 - factorized_top_k/top_100_categorical_accuracy: 0.3460 - loss: 22182.9654 - regularization_loss: 0.0000e+00 - total_loss: 22182.9654\n",
      "1/1 [==============================] - 3s 3s/step - factorized_top_k/top_1_categorical_accuracy: 0.0022 - factorized_top_k/top_5_categorical_accuracy: 0.0118 - factorized_top_k/top_10_categorical_accuracy: 0.0214 - factorized_top_k/top_50_categorical_accuracy: 0.0764 - factorized_top_k/top_100_categorical_accuracy: 0.1252 - loss: 15575.5381 - regularization_loss: 0.0000e+00 - total_loss: 15575.5381\n"
     ]
    }
   ],
   "source": [
    "train_results = model.evaluate(train_ds, return_dict=True)\n",
    "test_results = model.evaluate(test_ds, return_dict=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "id": "503ddc80",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train top-100 accuracy:  0.3460131585597992\n",
      "Test top-100 accuracy:  0.12521815299987793\n"
     ]
    }
   ],
   "source": [
    "print(f\"Train top-100 accuracy:  {train_results['factorized_top_k/top_100_categorical_accuracy']}\")\n",
    "print(f\"Test top-100 accuracy:  {test_results['factorized_top_k/top_100_categorical_accuracy']}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1fe3d884",
   "metadata": {},
   "source": [
    "There is still some overfitting, but it is much better than before and the top-100 accuracy is the best of all the models so far!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "id": "c11174bc",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Create a model that takes in raw query features, and\n",
    "index = tfrs.layers.factorized_top_k.BruteForce(model.user_model)\n",
    "# recommends items out of the entire items dataset.\n",
    "items_identifier_ds = items_dataset_w_context.map(lambda x: x['item_no'])\n",
    "_ = index.index_from_dataset(\n",
    "        tf.data.Dataset.zip((items_identifier_ds.batch(100), \n",
    "                             items_dataset_w_context.batch(100).map(model.item_model))))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "id": "2299d6a3",
   "metadata": {},
   "outputs": [],
   "source": [
    "random_user = np.random.choice(train_df_filtered['user_no'].unique())\n",
    "items_to_exclude = train_df_filtered.loc[train_df_filtered['user_no'] == random_user]['item_no'].unique()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "id": "f50c40d4",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 133 ms, sys: 4.58 ms, total: 137 ms\n",
      "Wall time: 184 ms\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "_, titles = index.query_with_exclusions(tf.constant([random_user]), \n",
    "                                       tf.constant([items_to_exclude]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "id": "d1dbeb56",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>user_no</th>\n",
       "      <th>item_no</th>\n",
       "      <th>gender_description</th>\n",
       "      <th>brand</th>\n",
       "      <th>product_group</th>\n",
       "      <th>first_interaction_month</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>15282</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>2727390070415090157</td>\n",
       "      <td>unisex</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>67539</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>941004878426231475</td>\n",
       "      <td>boys</td>\n",
       "      <td>champion</td>\n",
       "      <td>jumpers and knitwear</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>87417</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>-7150404021226002845</td>\n",
       "      <td>unisex</td>\n",
       "      <td>mini rodini</td>\n",
       "      <td>jumpers and knitwear</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>121109</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>-7679098013634930632</td>\n",
       "      <td>boys</td>\n",
       "      <td>timberland</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>130455</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>-4222212719177975921</td>\n",
       "      <td>unisex</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>158873</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>7902408375612206522</td>\n",
       "      <td>boys</td>\n",
       "      <td>champion</td>\n",
       "      <td>tops</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>175769</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>8398670320846353131</td>\n",
       "      <td>boys</td>\n",
       "      <td>moschino kid-teen</td>\n",
       "      <td>jumpers and knitwear</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>208072</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>7783670181111540799</td>\n",
       "      <td>girls</td>\n",
       "      <td>buddy &amp; hope</td>\n",
       "      <td>bottoms</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>208302</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>-826806050499523555</td>\n",
       "      <td>unisex</td>\n",
       "      <td>moschino kid-teen</td>\n",
       "      <td>jumpers and knitwear</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>241655</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>-2902137136503355060</td>\n",
       "      <td>boys</td>\n",
       "      <td>moncler</td>\n",
       "      <td>coats and jackets</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>292304</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>1625779193739963919</td>\n",
       "      <td>unisex</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>292560</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>-4467779472524472406</td>\n",
       "      <td>unisex</td>\n",
       "      <td>reima</td>\n",
       "      <td>scarves</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>294980</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>-508644785632326109</td>\n",
       "      <td>boys</td>\n",
       "      <td>stella mccartney kids</td>\n",
       "      <td>coats and jackets</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>321947</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>7260136798306039525</td>\n",
       "      <td>boys</td>\n",
       "      <td>champion</td>\n",
       "      <td>jumpers and knitwear</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>356542</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>-205346473013288548</td>\n",
       "      <td>boys</td>\n",
       "      <td>timberland</td>\n",
       "      <td>headwear</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>497346</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>8734965530751774714</td>\n",
       "      <td>unisex</td>\n",
       "      <td>kuling</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>523769</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>-6717261085387864616</td>\n",
       "      <td>unisex</td>\n",
       "      <td>moncler</td>\n",
       "      <td>coats and jackets</td>\n",
       "      <td>10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>532988</th>\n",
       "      <td>-355917100050929382</td>\n",
       "      <td>2992168747616124229</td>\n",
       "      <td>boys</td>\n",
       "      <td>champion</td>\n",
       "      <td>clothing sets</td>\n",
       "      <td>11</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                    user_no               item_no gender_description  \\\n",
       "15282   -355917100050929382   2727390070415090157             unisex   \n",
       "67539   -355917100050929382    941004878426231475               boys   \n",
       "87417   -355917100050929382  -7150404021226002845             unisex   \n",
       "121109  -355917100050929382  -7679098013634930632               boys   \n",
       "130455  -355917100050929382  -4222212719177975921             unisex   \n",
       "158873  -355917100050929382   7902408375612206522               boys   \n",
       "175769  -355917100050929382   8398670320846353131               boys   \n",
       "208072  -355917100050929382   7783670181111540799              girls   \n",
       "208302  -355917100050929382   -826806050499523555             unisex   \n",
       "241655  -355917100050929382  -2902137136503355060               boys   \n",
       "292304  -355917100050929382   1625779193739963919             unisex   \n",
       "292560  -355917100050929382  -4467779472524472406             unisex   \n",
       "294980  -355917100050929382   -508644785632326109               boys   \n",
       "321947  -355917100050929382   7260136798306039525               boys   \n",
       "356542  -355917100050929382   -205346473013288548               boys   \n",
       "497346  -355917100050929382   8734965530751774714             unisex   \n",
       "523769  -355917100050929382  -6717261085387864616             unisex   \n",
       "532988  -355917100050929382   2992168747616124229               boys   \n",
       "\n",
       "                        brand         product_group  first_interaction_month  \n",
       "15282                 småfolk           all in ones                       11  \n",
       "67539                champion  jumpers and knitwear                       11  \n",
       "87417             mini rodini  jumpers and knitwear                       11  \n",
       "121109             timberland           all in ones                       11  \n",
       "130455                småfolk           all in ones                       11  \n",
       "158873               champion                  tops                       11  \n",
       "175769      moschino kid-teen  jumpers and knitwear                       11  \n",
       "208072           buddy & hope               bottoms                       11  \n",
       "208302      moschino kid-teen  jumpers and knitwear                       11  \n",
       "241655                moncler     coats and jackets                       10  \n",
       "292304                småfolk           all in ones                       11  \n",
       "292560                  reima               scarves                       11  \n",
       "294980  stella mccartney kids     coats and jackets                       10  \n",
       "321947               champion  jumpers and knitwear                       11  \n",
       "356542             timberland              headwear                       11  \n",
       "497346                 kuling           all in ones                       11  \n",
       "523769                moncler     coats and jackets                       10  \n",
       "532988               champion         clothing sets                       11  "
      ]
     },
     "execution_count": 52,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_df_filtered.loc[train_df_filtered['user_no'] == random_user]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "id": "78d609cd",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>item_no</th>\n",
       "      <th>colour</th>\n",
       "      <th>gender_description</th>\n",
       "      <th>brand</th>\n",
       "      <th>product_group</th>\n",
       "      <th>min_age</th>\n",
       "      <th>max_age</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>1065</th>\n",
       "      <td>5310331093029041126</td>\n",
       "      <td>blue</td>\n",
       "      <td>boys</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>0.125</td>\n",
       "      <td>0.625</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>9680</th>\n",
       "      <td>-6590454116463224914</td>\n",
       "      <td>blue</td>\n",
       "      <td>boys</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>0.125</td>\n",
       "      <td>2.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>13531</th>\n",
       "      <td>-2692550025455852972</td>\n",
       "      <td>green</td>\n",
       "      <td>boys</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>0.125</td>\n",
       "      <td>2.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>13751</th>\n",
       "      <td>-6529328654082724751</td>\n",
       "      <td>blue</td>\n",
       "      <td>boys</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>0.125</td>\n",
       "      <td>2.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>22292</th>\n",
       "      <td>-3022787317072387909</td>\n",
       "      <td>green</td>\n",
       "      <td>boys</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>0.125</td>\n",
       "      <td>2.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>26803</th>\n",
       "      <td>1192638001384264749</td>\n",
       "      <td>yellow</td>\n",
       "      <td>boys</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>underwear</td>\n",
       "      <td>0.875</td>\n",
       "      <td>6.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>31310</th>\n",
       "      <td>3653303747819224244</td>\n",
       "      <td>blue</td>\n",
       "      <td>boys</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>0.125</td>\n",
       "      <td>2.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>46602</th>\n",
       "      <td>5338400809252936593</td>\n",
       "      <td>green</td>\n",
       "      <td>boys</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>0.125</td>\n",
       "      <td>2.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>50453</th>\n",
       "      <td>7554514181884221350</td>\n",
       "      <td>blue</td>\n",
       "      <td>boys</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>0.125</td>\n",
       "      <td>2.000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>60869</th>\n",
       "      <td>-5695807411571609566</td>\n",
       "      <td>blue</td>\n",
       "      <td>boys</td>\n",
       "      <td>småfolk</td>\n",
       "      <td>all in ones</td>\n",
       "      <td>0.125</td>\n",
       "      <td>0.625</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                    item_no  colour gender_description    brand product_group  \\\n",
       "1065    5310331093029041126    blue               boys  småfolk   all in ones   \n",
       "9680   -6590454116463224914    blue               boys  småfolk   all in ones   \n",
       "13531  -2692550025455852972   green               boys  småfolk   all in ones   \n",
       "13751  -6529328654082724751    blue               boys  småfolk   all in ones   \n",
       "22292  -3022787317072387909   green               boys  småfolk   all in ones   \n",
       "26803   1192638001384264749  yellow               boys  småfolk     underwear   \n",
       "31310   3653303747819224244    blue               boys  småfolk   all in ones   \n",
       "46602   5338400809252936593   green               boys  småfolk   all in ones   \n",
       "50453   7554514181884221350    blue               boys  småfolk   all in ones   \n",
       "60869  -5695807411571609566    blue               boys  småfolk   all in ones   \n",
       "\n",
       "       min_age  max_age  \n",
       "1065     0.125    0.625  \n",
       "9680     0.125    2.000  \n",
       "13531    0.125    2.000  \n",
       "13751    0.125    2.000  \n",
       "22292    0.125    2.000  \n",
       "26803    0.875    6.000  \n",
       "31310    0.125    2.000  \n",
       "46602    0.125    2.000  \n",
       "50453    0.125    2.000  \n",
       "60869    0.125    0.625  "
      ]
     },
     "execution_count": 53,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "recommendations = [item.numpy().decode() for item in titles[0]]\n",
    "item_info_df.loc[item_info_df['item_no'].isin(recommendations)]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1a303445",
   "metadata": {},
   "source": [
    "### **Performance**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "id": "d2c01c6d",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 626 ms, sys: 148 ms, total: 773 ms\n",
      "Wall time: 303 ms\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "_, retrieved_items = index(tf.constant(test_df_filtered['user_no']), k=100)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6ef64a45",
   "metadata": {},
   "source": [
    "Less than a second to predict the top 100 items for 2000 users, and that's with a brute force search "
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "tfrs-retail-example",
   "language": "python",
   "name": "tfrs-retail-example"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}