Title:Deoscillated Graph Collaborative Filtering

Abstract:Collaborative Filtering (CF) signals are crucial for a Recommender System~(RS) model to learn user and item embeddings. High-order information can alleviate the cold-start issue of CF-based methods, which is modelled through propagating the information over the user-item bipartite graph. Recent Graph Neural Networks~(GNNs) propose to stack multiple aggregation layers to propagate high-order signals. However, the oscillation problem, varying locality of bipartite graph, and the fix propagation pattern spoil the ability of multi-layer structure to propagate information. The oscillation problem results from the bipartite structure, as the information from users only propagates to items. Besides oscillation problem, varying locality suggests the density of nodes should be considered in the propagation process. Moreover, the layer-fixed propagation pattern introduces redundant information between layers. In order to tackle these problems, we propose a new RS model, named as \textbf{D}eoscillated \textbf{G}raph \textbf{C}ollaborative \textbf{F}iltering~(DGCF). We introduce cross-hop propagation layers in it to break the bipartite propagating structure, thus resolving the oscillation problem. Additionally, we design innovative locality-adaptive layers which adaptively propagate information. Stacking multiple cross-hop propagation layers and locality layers constitutes the DGCF model, which models high-order CF signals adaptively to the locality of nodes and layers. Extensive experiments on real-world datasets show the effectiveness of DGCF. Detailed analyses indicate that DGCF solves oscillation problem, adaptively learns local factor, and has layer-wise propagation pattern. Our code is available online at this https URL.