Reproduce “Contrastive Meta Learning with Behavior Multiplicity for Recommendation” on OpenResearch

About this paper

Wei Wei, Chao Huang, Lianghao Xia, Yong Xu, Jiashu Zhao, Dawei Yin

A Contrastive Meta Learning (CML) framework is introduced to enhance recommendation systems by modeling complex user-item relationships. It leverages a self-supervised multi-behavior contrastive learning paradigm and a meta-network to address sparse supervision signals and capture personalized multi-behavior patterns. The CML framework consistently outperforms state-of-the-art baselines across real-world datasets, notably improving HR@10 by 25.4% and NDCG@10 by 25.8% on the Tmall dataset.

Key insights

• Self-supervised contrastive learning across diverse user behavior views can effectively transfer knowledge and augment supervision signals, mitigating data sparsity for target behaviors.
• A meta-learning approach can capture and adaptively model personalized dependencies between different types of user behaviors, moving beyond a one-size-fits-all assumption.
• Combining a graph-based representation learning with personalized, weighted contrastive objectives significantly improves recommendation accuracy and provides interpretable insights into user behavior patterns.

Summary by alphaXiv.