Hierarchical fine-grained multi-behavior recommendation with behavior-aware contrastive learning.

Abstract

Real-life recommendation scenarios usually involve various interaction behaviors (e.g., click, cart, and purchase), carrying rich semantic information for learning multi-dimensional user preferences. Despite recent advances in leveraging multi-behavior interactions for recommendations, three key limitations persist: (i) Existing methods that ignore the varying influence of intents on different behaviors fail to infer true motivations behind each behavior, leading to irrelevant preference information under auxiliary behaviors being incorrectly transferred to target behavior. (ii) Prior GCN-based models struggle to effectively integrate high-order relations across both specific and global dimensions, leading to the loss of useful information. (iii) Current contrastive learning-based works attempt to maximize consistency across different behaviors to enable knowledge transfer, but perfect alignment between behavior embeddings destroys the uniqueness of each behavior. Thus, we propose the Hierarchical Fine-grained Multi-behavior Recommendation with Behavior-aware Contrastive Learning (HFCL) framework. Specifically, we first disentangle intents behind behaviors to learn users' fine-grained preferences and incorporate intent influence into multi-behavior semantics modeling. Next, we design dual-level graph convolutional networks to collaboratively leverage both types of high-order relations, thereby enhancing representation learning. Finally, we introduce a novel contrastive learning approach that ensures a better balance between consistency and uniqueness across different behaviors. Extensive experiments indicate that HFCL outperforms state-of-the-art baselines. Our implementation is accessible to the public at: https://github.com/zkyqust/HFCL.