This repository contains PyTorch codes and datasets for the paper:

Wei, Wei and Huang, Chao and Xia, Lianghao and Xu, Yong and Zhao, Jiashu and Yin, Dawei. Contrastive Meta Learning with Behavior Multiplicity forRecommendation. Paper in arXiv.

Contrastive Meta Learning (CML) leverages multi-behavior learning paradigm to model diverse and multiplex user-item relationships, as well as tackling the label scarcity problem for target behaviors. The designed multi-behavior contrastive task is to capture the transferable user-item relationships from multi-typed user behavior data heterogeneity. And the proposed meta contrastive encoding scheme allows CML to preserve the personalized multi-behavior characteristics, so as to be reflective of the diverse behavior-aware user preference under a customized self-supervised framework.

@inproceedings{wei2022contrastive,
  title={Contrastive meta learning with behavior multiplicity for recommendation},
  author={Wei, Wei and Huang, Chao and Xia, Lianghao and Xu, Yong and Zhao, Jiashu and Yin, Dawei},
  booktitle={Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining},
  pages={1120--1128},
  year={2022}
}

The codes of CML are implemented and tested under the following development environment:

• Python 3.6
• torch==1.8.1+cu111
• scipy==1.6.2
• tqdm==4.61.2

• IJCAI contest: https://tianchi.aliyun.com/dataset/dataDetail?dataId=47
• Retail Rocket: https://www.kaggle.com/retailrocket/ecommerce-dataset
• Tmall: https://tianchi.aliyun.com/dataset/dataDetail?dataId=649

• The processed IJCAI are under the /datasets folder.

The command to train CML on the Tmall/IJCAI/Retailrocket datasets are as follows. The commands specify the hyperparameter settings that generate the reported results in the paper.

• Tmall

python main.py --path=./datasets/ --dataset=Tmall --opt_base_lr=1e-3 --opt_max_lr=5e-3 --opt_weight_decay=1e-4 --meta_opt_base_lr=1e-4 --meta_opt_max_lr=2e-3 --meta_opt_weight_decay=1e-4 --meta_lr=1e-3 --batch=8192 --meta_batch=128 --SSL_batch=18

• IJCAI

python main.py --path=./datasets/ --dataset=IJCAI_15 --sampNum=10 --opt_base_lr=1e-3 --opt_max_lr=2e-3 --opt_weight_decay=1e-4 --meta_opt_base_lr=1e-4 --meta_opt_max_lr=1e-3 --meta_opt_weight_decay=1e-4 --meta_lr=1e-3 --batch=8192 --meta_batch=128 --SSL_batch=30 

• Retailrocket

python main.py --path=./datasets/ --dataset='retailrocket' --sampNum=40 --lr=3e-4 --opt_base_lr=1e-4 --opt_max_lr=1e-3 --opt_weight_decay=1e-4 --opt_weight_decay=1e-4 --meta_opt_base_lr=1e-4 --meta_opt_max_lr=1e-3 --meta_opt_weight_decay=1e-3 --meta_lr=1e-3 --batch=2048 --meta_batch=128 --SSL_batch=15

It will be released again in few days in the optimized code version.