jaywonchung / BERT4Rec-VAE-Pytorch

Pytorch implementation of BERT4Rec and Netflix VAE.
GNU General Public License v3.0
356 stars 84 forks source link
ae bert4rec dae pytorch recommendation-system vae

Introduction

This repository implements models from the following two papers:

BERT4Rec: Sequential Recommendation with BERT (Sun et al.)

Variational Autoencoders for Collaborative Filtering (Liang et al.)

and lets you train them on MovieLens-1m and MovieLens-20m.

Usage

Overall

Run main.py with arguments to train and/or test you model. There are predefined templates for all models.

On running main.py, it asks you whether to train on MovieLens-1m or MovieLens-20m. (Enter 1 or 20)

After training, it also asks you whether to run test set evaluation on the trained model. (Enter y or n)

BERT4Rec

python main.py --template train_bert

DAE

python main.py --template train_dae

VAE

Search for the optimal beta

python main.py --template train_vae_search_beta

Use the found optimal beta

First, fill out the optimal beta value in templates.py. Then, run the following.

python main.py --template train_vae_give_beta

The Best_beta plot will help you determine the optimal beta value. It can be seen that the optimal beta value is 0.285.

The gray graph in the Beta plot was trained by fixing the beta value to 0.285.

The NDCG_10 metric shows that the improvement claimed by the paper has been reproduced.

Examples

  1. Train BERT4Rec on ML-20m and run test set inference after training

    printf '20\ny\n' | python main.py --template train_bert
  2. Search for optimal beta for VAE on ML-1m and do not run test set inference

    printf '1\nn\n' | python main.py --template train_vae_search_beta

Test Set Results

Numbers under model names indicate the number of hidden layers.

MovieLens-1m

MovieLens-20m