Fine-Grained Representation Learning and Multi-View Collaborative Augmentation for Recommendation

Abstract

Graph neural networks (GNNs) have recently advanced in processing graph-structured data and are increasingly used in recommendation systems. Recently, many studies have incorporated side information as auxiliary views, such as the user’s social connections and the item’s knowledge-aware dependencies, to enhance the user-item interaction view. However, current works overlook the differences in learning behavior between auxiliary views and interaction view, and transfer side information from different views separately, which can lead to a semantic gap and fail to explore the collaborative effect of auxiliary views. To address this challenge, we propose FiCoRec, a novel fine-grained augmentation method for recommendation, comprising two key enhancement components: Hierarchical Knowledge Transfer (HKT) and Multi-view Semantic Fusion (MSF). Specifically, HKT designs an interaction semantic decouple (ISD) method to decouple the interaction view embeddings into homogeneous features (hoFs) and heterogeneous features (heFs). Then a hierarchical contrastive learning framework is used to fully capture the local and global semantics from the intermediate-layer to enhance hoFs. MSF explores a collaborative augmentation mechanism by utilizing meta-learning to enhance the interaction view. Extensive experiments conducted on five datasets against seven baseline methods demonstrate that our FiCoRec outperforms the state-of-the-art methods with a margin of 0.33%–2.76%.