Interaction Intent-Aware Graph Contrastive Learning for Recommendation

In the realm of recommender systems, leveraging graph contrastive learning (GCL) for addressing issues with scarce supervised signals has emerged as a prominent trend, demonstrating impressive performance improvements through augmented data views and maximized consistency between them. However, most GCL-based methods rely simply on binary user-item feedback to construct CL tasks, lacking in the exploration of nuanced preference levels underlying user interactions, thereby leading to recommendations misaligned with users' actual intents. To address these problems, we introduce Interaction Intent-Aware Graph Contrastive Learning (IACL), a method specifically designed to integrate intent modeling with contrastive learning for recommendation. We propose to model the preference levels behind user interaction behaviors by learning interaction weights from multiple aspects, followed by fusing these weights to create a weighted interaction view. Then, we innovatively utilize the fused weights to decouple user intents by associating each interaction with an intent vector. Building upon this, intent-aware and weight-enhanced dual contrastive learning tasks are designed to optimize intent modeling and refine user and item representations. Extensive experiments conducted on three public datasets demonstrate the effectiveness of the proposed IACL.