Deep Multifaceted Transformers for Multi-objective Ranking in Large-Scale E-commerce Recommender Systems (CIKM 2020)

DMT:

DMT_code is the code for the paper "Deep Multifaceted Transformers for Multi-objective Ranking in Large-Scale E-commerce Recommender Systems", which is published in CIKM 2020.

Requirements

python==2.7

tensorflow==1.12

Usage

sh run.sh

Datasets:

JD Recsys Dataset.

Statistics:

Type	Total Sampled	Impressions	Clicks	Orders
Train	667,907,650	622,596,211	43,876,602	1,434,837
Test	105,444,671	98,732,799	6,477,409	234,463

Download link:

The dataset can be downloaded from: https://drive.google.com/drive/folders/1Dnlnnzl2QD2mYP3o0icSxNVrl6nCvlT0?usp=sharing. The files are in the format of TFRecord and can be placed in HDFS for training.

The shared dataset is sampled from the 0.7 billion dataset used in the paper.

Description:

The datasets are used in "Deep Multifaceted Transformers for Multi-objective Ranking in Large-Scale E-commerce Recommender Systems", which is published in CIKM 2020.

In this paper, the two tasks are click prediction and order prediction.

The goal of click prediction is to predict the CTR.

The goal of order prediction is to predict the CTVR = CTR * CVR, which aims to eliminate the sample bias problem in CVR task [1] .

CTR: Click-Through Rate, CVR: ConVersion Rate.

Type	Click prediction (CTR)	Order prediction (CTVR)
Impression	0	0
Click	1	0
Order	1	1

Research Topics:

This dataset can be used for research on CTR prediction, CVR prediction, multi-task ranking, sequential modeling, unbiased ranking in Recommender Systems. This dataset should only be used for research purpose!

Citation:

Please cite the following paper if you use the data in any way.

@inproceedings{gu2020dmt,
  title={Deep Multifaceted Transformers for Multi-objective Ranking in Large-Scale E-commerce Recommender Systems},
  author={Gu, Yulong and Ding, Zhuoye and Wang, Shuaiqiang and Zou, Lixin and Yiding Liu and Yin, Dawei},
  booktitle={CIKM'20},
  year={2020}
}

File Description:

The dataset files are in the format of TFRecord.

Each line in the TFRecord files contains the ranking features and label in the Recommender System.

The ranking features contains 615 dense features and some id features.

The labels are impressions (labels=0), clicks (labels=1 or 2) and orders (labels=4 or 5).

feature type	name	desc
Dense features	features	615 dimension dense features. They contain item profile features (e.g.,number of clicks, CTR, CVR, rating) , use profile features (e.g.,prefered categories and brands, purchase power), user-item matching features (e.g.,whether the item matches the user’s gender or age) and user-item interaction features (e.g.,number of clicks on thecategory of the item within a time window).
Categorical features	item_fea_sku	id of the product
	item_c2	second level category id of the product
	item_c3	third level category id of the product
	item_brand	brand id of the product
	item_shop	shop id of the product
clk_seq	clk_seq_sku_7d_50	sequence of ids of the products in the click sequence (latest 50 clicks in recent 7 days)
	clk_seq_ts_7d_50	sequence of timestamps in the click sequence (latest 50 clicks in recent 7 days)
	clk_seq_c2_7d_50	sequence of second level category ids in the click sequence (latest 50 clicks in recent 7 days)
	clk_seq_c3_7d_50	sequence of third level category ids in the click sequence (latest 50 clicks in recent 7 days)
	clk_seq_brand_7d_50	sequence of brand ids in the click sequence (latest 50 clicks in recent 7 days)
	clk_seq_shop_7d_50	sequence of shop ids in the click sequence (latest 50 clicks in recent 7 days)
ord_seq	ord_seq_sku_12m_50	sequence of ids of the products in the purchase sequence (latest 50 orders in recent 12 months)
	ord_seq_ts_12m_50	sequence of timestamps in the purchase sequence (latest 50 orders in recent 12 months)
	ord_seq_c2_12m_50	sequence of second level category ids in the purchase sequence (latest 50 orders in recent 12 months)
	ord_seq_c3_12m_50	sequence of third level category ids in the purchase sequence (latest 50 orders in recent 12 months)
	ord_seq_brand_12m_50	sequence of brand ids in the purchase sequence (latest 50 orders in recent 12 months)
	ord_seq_shop_12m_50	sequence of shop ids in the purchase sequence (latest 50 orders in recent 12 months)
cart_seq	cart_seq_sku_12m_10	sequence of ids of the products in the cart sequence (latest 10 carts in recent 12 months)
	cart_seq_ts_12m_10	sequence of timestamps in the cart sequence (latest 10 carts in recent 12 months)
	cart_seq_c2_12m_10	sequence of second level category ids in the cart sequence (latest 10 carts in recent 12 months)
	cart_seq_c3_12m_10	sequence of third level category ids in the cart sequence (latest 10 carts in recent 12 months)
	cart_seq_brand_12m_10	sequence of brand ids in the cart sequence (latest 10 carts in recent 12 months)
	cart_seq_shop_12m_10	sequence of shop ids in the cart sequence (latest 10 carts in recent 12 months)
Bias features	near_expo_seq_c2	sequence of second level category ids of the neighboring exposured products of the product
	near_expo_seq_c3	sequence of third level category ids of the neighboring exposured products of the product
	page	page number of the product exposed in the page
	position	position number of the product exposed in the page

References:

[1] Ma, Xiao, Liqin Zhao, Guan Huang, Zhi Wang, Zelin Hu, Xiaoqiang Zhu, and Kun Gai. "Entire space multi-task model: An effective approach for estimating post-click conversion rate." SIGIR'2018.