minzc/torchease

Implementation of Embarrassingly Shallow Autoencoders (Harald Steck) in PyTorch

TorchEASE

Implementation of Embarrassingly Shallow Autoencoders (Harald Steck) in PyTorch

* Adapted from: https://github.com/Darel13712/ease_rec 
* Paper: https://arxiv.org/abs/1905.03375
* Papers with Code: https://paperswithcode.com/sota/collaborative-filtering-on-million-song

This code is currently SOTA on the Million Song challenge on Papers With Code. It utilizes a simple closed-form solution for a recommendation problem. This PyTorch implementation is much faster than the original Numpy version.

To use:

Implicit interactions:

1. te_implicit = TorchEASE(df, user_col="user", item_col="item")
2. te_implicit.fit()
3. predictions = te_implicity.predict_all(predict_df)

Explicit interactions (UNTESTED):

1. te_explicit = TorchEASE(df, user_col="user", item_col="item", score_col="rating")
2. te_explicit.fit()
3. predictions = te_explicit.predict_all(predict_df)

To use via CLI:

1. Put your data files into data/
2. From a terminal window, call python train.py {TRAIN.csv} {TO_PRED.csv} {USER_COLUMN_NAME} {ITEM_COLUMN_NAME} {[optional]SCORE_COLUMN_NAME}
3. Predictions are pushed to predictions/output.csv

Example usage:

python train.py training.csv users.csv username book_name book_score

Tuning the regularization parameter on scales from 1E2 - 1E3 seems to be effective