Multi-modality Meets Re-learning: Mitigating Negative Transfer in Sequential Recommendation

This is the implementation of the $\mathtt{ANT}$ model in our paper (https://arxiv.org/abs/2309.10195).

Dependency

• python==3.9.13
• torch==1.10.2
• pytorch_lightning==1.7.7
• recbole==1.1.1

We encourage users to install the dependencies using pip install

Processed Dataset

Our data processing scripts are built upon the scripts of UniSRec. Due to the copyright consideration, we realse the processed dataset instead of the processing scripts. If you are interested in the scripts, please send a request to Bo Peng (peng.707@osu.edu).

For the datasets used in evaluation, due to the space limit in GitHub, we release the processed data in Google Drive. Please download the processed data and put it under the "data/SWIN_BERT/" folder to run the training scripts.

The data used for pre-training is too large even for Google Drive. Please contact Bo Peng directly if you are interested.

Pre-trained Recommendation Model

We release our pre-trained recommendation model in the folder "best_model/pretrain/".

Adapting the Pre-trained Model

Please refer to the following example on how to adapt the pre-trained recommendation model on a target task (e.g., Scientific).

python finetune.py --datasets=Scientific --fusion_lam=0.1 --max_epochs=100 --pretrained_model=Clothing_Home_Movies_Food_0.1_0_0

fusion_lam specifies the coefficien for modality fusion $\beta$.

Model Pre-training

Please refer to the following example on how to pretrain the recommendation model

python pretrain.py --datasets=Clothing_Home_Movies_Food --training_stage=pretrain --max_epochs=300

Acknowledgements

Our data processing scripts are built on the scripts of UniSRec. Thanks for the great work!