RecDCL: Dual Contrastive Learning for Recommendation
WWW 2024(2024)
摘要
Self-supervised recommendation (SSR) has achieved great success in mining the
potential interacted behaviors for collaborative filtering in recent years. As
a major branch, Contrastive Learning (CL) based SSR conquers data sparsity in
Web platforms by contrasting the embedding between raw data and augmented data.
However, existing CL-based SSR methods mostly focus on contrasting in a
batch-wise way, failing to exploit potential regularity in the feature-wise
dimension, leading to redundant solutions during the representation learning
process of users (items) from Websites. Furthermore, the joint benefits of
utilizing both Batch-wise CL (BCL) and Feature-wise CL (FCL) for
recommendations remain underexplored. To address these issues, we investigate
the relationship of objectives between BCL and FCL. Our study suggests a
cooperative benefit of employing both methods, as evidenced from theoretical
and experimental perspectives. Based on these insights, we propose a dual CL
method for recommendation, referred to as RecDCL. RecDCL first eliminates
redundant solutions on user-item positive pairs in a feature-wise manner. It
then optimizes the uniform distributions within users and items using a
polynomial kernel from an FCL perspective. Finally, it generates contrastive
embedding on output vectors in a batch-wise objective. We conduct experiments
on four widely-used benchmarks and an industrial dataset. The results
consistently demonstrate that the proposed RecDCL outperforms the
state-of-the-art GNNs-based and SSL-based models (with up to a 5.65\%
improvement in terms of Recall@20), thereby confirming the effectiveness of the
joint-wise objective. All source codes used in this paper are publicly
available at \url{https://github.com/THUDM/RecDCL}}.
更多查看译文
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要