【摘要】 基于图卷积网络(GCN)模型在学习用户/物品表示方面表现出了强大的性能,给传统的协作过滤(CF)算法带来了新的研究突破。然而,现有的基于GCN的CF方法仍然都是针对静态图建模,而在实际场景中,用户与物品的交互不是一成不变的,会随着时间的推移而持续演化;GCN中的过平滑问题会极大地限制现有推荐算法的表示学习建模。为解决上述问题,提出了基于动态图的协同过滤算法(DynGCF),其目的是通过同时捕获图的结构和时态演化信息来学习用户和物品的嵌入表示。DynGCF首先采用GCN学习每个离散快照图上的用户/物品嵌入,然后应用时间卷积网络(TCN)和自注意力机制学习,最终嵌入表示。为缓解过平滑问题,本文改进了传统GCN中的关键模块,即邻域聚合,通过在1阶交互图和2阶共现图建模用户和物品的交互。在4个真实数据集上与基于GCN的CF方法和动态图的基线方法对比,验证了DynGCF的性能提升,并分析验证了改进的方法能有效缓解过平滑问题。更多还原

【Abstract】 The graph convolution network(GCN) based models have shown powerful performance in learning the users’/items’ representations and achieved new state-of-the-art for collaborative filtering( CF). Nevertheless, existing GCN based CF methods still have following limitations: They are all target static graphs while many real-life graphs evolve over time. Existing work that adapts GCN to recommender systems suffers from performance limitations due to the over-smoothing issue. To tackle the aforementioned problems, a dynamic graph based collaborative filtering(DynGCF) is proposed, which aims to learn the representations of users and items by capturing both graph structural and temporal information. Specifically, DynGCF adapts GCN to learn discrete user/item embeddings on each graph snapshot at first, then employs temporal convolutional networks(TCN) and self-attention mechanism to learn the final embeddings. To alleviate the over-smoothing issue, we analyze and simplify the neighborhood aggregation,which is a pivot component in GCN, by jointly using only 1-hop interaction and 2-hop co-occurrence graph to model the user-item interactions. Extensive experiments are conducted on four real-world datasets to demonstrate the significant performance gains for DynGCF over state-of-the-art GCN based CF methods and dynamic graph based methods. Further analysis proves that the alleviation of the over-smoothing benefits from the hop graphs.更多还原