Skip to content

Latest commit

 

History

History
149 lines (122 loc) · 7.4 KB

README.md

File metadata and controls

149 lines (122 loc) · 7.4 KB
Raw

Towards Explainable Real Estate Valuation via Evolutionary Algorithms - Code

See https://arxiv.org/abs/2110.05116 for the paper this code belongs to.

We know this is scientific code and, at best, doubiously documented. We are happy to help, contact us at [email protected].

What is this?

This repo contains scientific code to evaluate the algorithms presentend in the paper experimentally.

The two parts

The paper compares our CBR+EA approach to more traditional DNNs. The folder rust-db contains code to execute the CBR+EA algorithm and python-dbcontains code to execute the comparision DNNs.

The Database

The algorithms work with training and testing data which is loaded from a Mongo DB. You have to setup a MongoDB which contains your dataset in a collection in the following schema. Then place the connection string into the environment variable MONGO_HOST in the following format

export MONGO_HOST="export MONGO_HOST="mongodb://[user]:[encoded pw]@[hostname]:[port]/[database_name]"

If your dataset has a different schema, you mainly need to adjust python-db/data/data.py and rust-db/common/src/immo.rs to match it. In Rust, the compiler will guide you to adjust all other relevant places as well.

Schema

(Note that we load more data, in both code bases but only use those listed here for prediction.)

column name interpretation
marktwert asking price
kurzgutachten.objektangabenBaujahr year of construction
kurzgutachten.objektangabenWohnflaeche living area
grundstuecksgroesseInQuadratmetern lot area
average_loca average local price
balcony_area balcony area
wertermittlungsstichtag offer date
walk_distance1 distance to public transport
convenience_distance distance to convenicence store
school_ele_distance distance to elementary school
school_jun_distance distance to junior high school
distance_parking distance to parking
house_kaisuu floor
plane_location location (coordinate tuple)
objektunterart object type
land_toshi urbanity score

Running the CBR+EA

Installation

You need a recent rust compiler with cargo. We recommend installation by rustup. You additonally need openssl, pkgconfig and dhall.

Running

Note, place commandline arguments passed to the algorithm after the double dash --. Arguments before the double dash are passed to the Rust compiler.

From within the rust-db folder, run

$ cargo run -p executables --bin evo --release -- --help

USAGE:
    evo [FLAGS] [OPTIONS]

FLAGS:
        --evaluate-all-generations-on-test-set    Evaluate the best individual of each generation the test set
        --expert
    -h, --help                                    Prints help information
        --local-search
        --median
            Use weighted median instead of weighted average prediction. Default: false

        --unclean                                 Don't perform data cleaning
    -V, --version                                 Prints version information

OPTIONS:
    -c, --collection <collection>                   [default: cleaned_80]
        --config <config-path>
    -l <limit>
    -o, --output <output-path>
    -q, --query <query-path>
    -s, --seed <seed-path>
        --split-at-date <split-at-date>
            Split train/validation data at a given date. Enter in YYYY-MM-DD format.

        --split-by-hash-key <split-by-hash-key>
            Split into training and test set deterministically based on hashing an object with following key

    -w, --weight-output <weight-output-path>
        --write-to-mongo <write-to-mongo>          Write predictions to mongodb. Argument: key to use in the database.

Running the comparison DNNs

Installation

Install dependencies via pip:

pip install -r python-db/requirements.txt

Run the DNNs

From within python-db/machine_learning/deep_neural_networks, run

$ python main.py --help

usage: main.py [-h] [-l [LIMIT]] [-c [COLLECTION]] [-n [NETWORK]] [-is [INPUT_SCALER]] [-os [OUTPUT_SCALER]]
               [-sp [TRAIN_TEST_SPLIT_DATE]] [-k [K_FOLD]] [--predefined PREDEFINED] [--epochs EPOCHS]
               [--network-from-pickle NETWORK_FROM_PICKLE]
               [--network-from-pickle-nth-best NETWORK_FROM_PICKLE_NTH_BEST] [--unclean-japan]
               [--write-to-mongo WRITE_TO_MONGO] [--sort-by-date] [--sort-by-hash SORT_BY_HASH]
               [--cbr-column CBR_COLUMN] [--tensorflow-random-seed TENSORFLOW_RANDOM_SEED]

Train and score a DNN on real estate data from germany

optional arguments:
  -h, --help            show this help message and exit
  -l [LIMIT], --limit [LIMIT]
                        set the limit for reading data
  -c [COLLECTION], --collection [COLLECTION]
                        set the collection to use
  -n [NETWORK], --network [NETWORK]
                        set the network to use
  -is [INPUT_SCALER], --input-scaler [INPUT_SCALER]
                        set the input scaler to use
  -os [OUTPUT_SCALER], --output-scaler [OUTPUT_SCALER]
                        set the output scaler to use
  -sp [TRAIN_TEST_SPLIT_DATE], --train-test-split-date [TRAIN_TEST_SPLIT_DATE]
                        Set the splitting date. Everything before is training-data, after is test data. Format:
                        yyyy-mm-dd
  -k [K_FOLD], --k-fold [K_FOLD]
                        turn on k-fold cross validation
  --predefined PREDEFINED
                        use tabnet/...
  --epochs EPOCHS       max number of epoch to train
  --network-from-pickle NETWORK_FROM_PICKLE
                        Load network from pickle. Set --network-from-pickle-nth-best to not use the best
  --network-from-pickle-nth-best NETWORK_FROM_PICKLE_NTH_BEST
                        See --network-from-pickle
  --unclean-japan       Dont filter by marktwert. Don't remove none values
  --write-to-mongo WRITE_TO_MONGO
                        write predictions to column in mongo collection
  --sort-by-date        sort all immos by date first. Doing this makes the train-val-split a "into-the-
                        future"-split
  --sort-by-hash SORT_BY_HASH
                        <seed>. Sort all immos by deterministic hash with seed. useful for comparable random
                        split
  --cbr-column CBR_COLUMN
                        read cbr results from column in mongo
  --tensorflow-random-seed TENSORFLOW_RANDOM_SEED
                        Sets the seed used for initalization of random weights

Notes

The file mongo/japan_recode_sales.json contains code to transform a collection as read from the data provided by the Japanese National Institute of Informatics into a collection format our code can work with.