Hierarchical Graph Contrastive Learning for Review-Enhanced Recommendation

Abstract

In comparison to numerical ratings and implicit feedback, textual reviews offer a deeper understanding of user preferences and item attributes. Recent research underscores the potential of reviews in augmenting recommendation capabilities, thereby advancing the deployment of review-enhanced recommendation systems. However, existing methodologies often neglect the significance of rating magnitudes and are susceptible to challenges such as data sparsity and long-tail distribution in real-world contexts. To address these challenges, we propose Hierarchical Graph Contrastive Learning (HGCL) for advancing review-enhanced recommendation systems. HGCL dynamically learns hypergraph structures to capture higher-order correlations among nodes and simultaneously integrates local and global collaborative relations through global-local contrastive learning. Additionally, we propose hierarchical graph contrastive learning methods to better model the intrinsic correlation between ratings and reviews, encompassing aspects such as local-global, cross-rating, and edge-level contrastive learning. Extensive experimentation on five public datasets demonstrates that the proposed method notably outperforms state-of-the-art approaches.