Graph-ICF: Item-based collaborative filtering based on graph neural network

摘要

Item-based collaborative filtering (ICF) has been widely used in industrial applications due to its good interpretability and flexible composability. The main idea of ICF is to characterize users’ preferences and recommend similar items based on the items interacted by users in the history logs. As such, the key problem for ICF approaches is how to learn item–item similarity effectively. Existing ICF approaches in general obtain an item’s feature by mapping from the embedding matrix, where only the shallow relation between the target item and the user’s historical interactive items can be captured, while the deeper relations conveyed by other users similar to the current user are neglected. To tackle this problem, we propose a graph-based ICF method (Graph-ICF), which takes advantage of the information aggregation and propagation properties of graph structure to dig out the deeper item relations that have been overlooked by existing ICF models. Moreover, existing ICF models only utilize item-level attention to discriminate which item–item similarities are important for a prediction, without considering user’s personalized favor of interacting with an item. To address this limitation, we further propose a feature level attention module in Graph-ICF to distinguish which feature dimensions better reflect the user’s preferences, and thus can yield more personalized recommendation. We conduct extensive experiments on three public benchmarks, demonstrating the superior performance of Graph-ICF over several state-of-the-art ICF models and graph-based CF methods.