Intent-guided Heterogeneous Graph Contrastive Learning for Recommendation

TL;DR

This work tackles the challenge of sparse interactions and latent intents in heterogeneous graphs for recommendation. It introduces Intent-guided Heterogeneous Graph Contrastive Learning (IHGCL), which combines a meta-path based Dual Contrastive Learning (DCL) module with a Bottlenecked AutoEncoder (BAE) to capture and denoise intents implicit in meta-paths. The approach jointly optimizes a BPR loss with information-bottleneck regularized denoising and dual contrastive losses, demonstrating state-of-the-art performance across six real-world datasets while offering insights into meta-path selection and robustness to noise. The results highlight the practical impact of explicitly modeling user and item intents via meta-paths in HGs to enhance contrastive learning for recommendations, with code available at the project repository.

Abstract

Contrastive Learning (CL)-based recommender systems have gained prominence in the context of Heterogeneous Graph (HG) due to their capacity to enhance the consistency of representations across different views. However, existing frameworks often neglect the fact that user-item interactions within HG are governed by diverse latent intents (e.g., brand preferences or demographic characteristics of item audiences), which are pivotal in capturing fine-grained relations. The exploration of these underlying intents, particularly through the lens of meta-paths in HGs, presents us with two principal challenges: i) How to integrate CL with intents; ii) How to mitigate noise from meta-path-driven intents. To address these challenges, we propose an innovative framework termed Intent-guided Heterogeneous Graph Contrastive Learning (IHGCL), which designed to enhance CL-based recommendation by capturing the intents contained within meta-paths. Specifically, the IHGCL framework includes: i) a meta-path-based Dual Contrastive Learning (DCL) approach to effectively integrate intents into the recommendation, constructing intent-intent contrast and intent-interaction contrast; ii) a Bottlenecked AutoEncoder (BAE) that combines mask propagation with the information bottleneck principle to significantly reduce noise perturbations introduced by meta-paths. Empirical evaluations conducted across six distinct datasets demonstrate the superior performance of our IHGCL framework relative to conventional baseline methods. Our model implementation is available at https://github.com/wangyu0627/IHGCL.

Figures (11)

• Figure 1: (a) Interaction graph in a movie scenario, where the green arrow indicates recommending movie $M_2$ to user $U_2$, and the red arrow indicates user $U_3$ to movie $M_3$; (b) heterogeneous graph incorporating user and item intents, showing that interactions are guided by intents; (c) considering the intents driven by the meta-paths.
• Figure 2: The impact w.r.t. different meta-paths. The blue bar is the number of added edges current meta-path, and the green line indicates the performance for current meta-path.
• Figure 3: A toy example of heterogeneous graph on Movielens zhang2023revisiting dataset for recommendation.
• Figure 4: The complete framework of the proposed IHGCL, which consists of a Dual Contrastive Learning (DCL) module and a Bottlenecked Autoencoder (BAE). The DCL module generates two types of contrasts: contrasts between meta-paths and contrasts between meta-path-enhanced views, providing self-supervised signals. The BAE module employs a dual-masked autoencoder combined with an adaptive information bottleneck technique to mitigate the noise issues, which can capture the minimum sufficient information from the data features.
• Figure 5: Theoretical analysis of DCL and BAE.