Group recommendation fueled by noise-based graph contrastive learning

Abstract

The ongoing advancement of social network platforms has increased the frequency of group activities. Due to the varied composition of group members, recommending items that align with the preferences of the entire group becomes a challenge. Existing group recommendations primarily deduce the final group decision by dynamically aggregating the preferences of group members. Although group recommendation involves various interaction types, the supervised signal sparsity problem caused by data sparsity limits the effectiveness of supervised learning, resulting in group preference representation suboptimal. To address this challenge, we propose a novel group recommendation (GR) approach named XCL-GR, which utilizes XsimGCL’s cross-layer contrastive learning (CL) for enhancement. Firstly, this approach leverages LightGCN’s characteristic of employing the initial embedding as the only update parameter to achieve simultaneous learning of multiple subgraphs. Secondly, data augmentation is achieved by introducing controlled noise into the embedding space to address the challenge of multi-interaction sparsity. Thirdly, this study employs a combined approach of supervised learning and self-supervised learning to address the challenge of sparse supervision signals across multiple learning tasks in group recommendation. The experiment was conducted on two real-world datasets, and the results demonstrate a significant improvement in the performance of both group and user recommendations.

Keywords

• Group recommendation
• Graph neural network
• Self-supervised learning
• Contrastive learning