Introduction

This is our PyTorch implementation for the paper:

MESA: Meta-model Enhanced Self-Attentive Sequential Recommendation

It includes the codes for BERT4Rec+MESA on two datasets (Movielens-1M and Dbook). You can change '/models' directory to use other base models.

Please cite our paper if you use the code.

The codes are based on the PyTorch implementation of SASRec and BERT4Rec.

https://github.com/pmixer/SASRec.pytorch

https://github.com/jaywonchung/BERT4Rec-VAE-Pytorch

Environment

PyTorch 1.12.0

Python 3.9.7

Datasets

The preprocessed datasets are included in the repo (ml-1m/data/ml-1m, dbook/data/dbook)

Model Training & Testing

You can change the --modulation_mode argument between 'shifting' and 'scaling' to use different modulation methods of MESA.

Movielens-1M

First, enter 'ml-1m' directory.

Training:

python main.py --dataset=ml-1m --train_dir=default --device cuda --template train_bert --modulation_mode shifting

The trained model will be saved as 'ml-1m_default/?.pth'. Find its name and replace the '?'.

Testing:

python main.py --device=cuda --dataset=ml-1m --train_dir=default --state_dict_path='ml-1m_default/?.pth' --inference_only=true --template train_bert --modulation_mode shifting

Dbook

First, enter 'dbook' directory.

Training:

python main.py --dataset=dbook --train_dir=default --device cuda --template train_bert --modulation_mode shifting

The trained model will be saved as 'dbook_default/?.pth'. Find its name and replace the '?'.

Testing:

python main.py --device=cuda --dataset=dbook --train_dir=default --state_dict_path='dbook_default/?.pth' --inference_only=true --template train_bert --modulation_mode shifting