Improving Temporal Graph Collaborative Filtering with Temporal Local Augmentation

Modeling the temporal evolution of users' preferences is essential for continuous-time sequential recommendation. Recently, temporal graph collaborative filtering (TGCF) applies temporal graph neural networks on user-item temporal graphs to capture the dynamic evolution. Despite the remarkable success, existing TGCF methods still suffer from long-tail issues, which can be described in two parts: (1) the long-tailed distribution of the node degrees in user-item temporal graphs; (2) the uncertainties when modeling the evolution on user-item temporal graphs. However, few temporal graph-based recommenders have studied how to alleviate the long-tail issues. In this paper, to tackle long-tail challenges, we introduce a novel approach: Temporal Local Augmentation (TLA) for temporal graph collaborative filtering. This method addresses the distribution and evolution of topology and features separately. Initially, we propose a pretraining phase to capture one-order node feature distributions using a generative model. Subsequently, we decouple the pretraining phase from downstream TGCF training and integrate memory with generated augmentation node features to bolster long-tail nodes. Experimental findings on five real-world datasets showcase TLA's effectiveness in mitigating long-tail challenges.