Research On Item-based Collaborative Filtering Based On Graph Neural Network

With the explosive growth of commodity quantities and types,personalized recommendation,which can predict the purchase intention of users,has become one of the most important services on the Internet.The personalized recommendation aims to predict a set of items that users are more likely to purchase in the future by fully employing users' historical interactions.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.Existing ICF approaches still have limitations,such as,when ICF approaches predict a user's preference,they rely more on the information of the user's historical interactive items,while not fully taking account of the information passed by similar users and silimar users' historical interactive items.As such,it may lead to the problem of not effectively capturing the collaborative filtering effect.It is a good solution that through expressive modeling the high-order connectivity in user-item graph,we effectively inject the collaborative signal into the embedding process(mapping from pre-existing features that describe the item,such as ID and attributes)in an explicit manner.To tackle this problem,we propose a graph-based ICF method,named Graph-ICF,which takes convolutional operations in user-item interaction graph,and takes advantage of the information aggregation and propagation properties of graph structure.Graph-ICF explicitly encodes the crucial collaborative signal which is latent in user-item interactions to effectively capturing the collaborative filtering effect.Moreover,we further propose a feature level attention module in Graph-ICF to distinguish which feature dimensions reflect the user's preferences.For the multiple embeddings obtained by feature propagating in GCN,we explore two ways of residual connection and weighted sum.We conduct extensive experiments on three public benchmarks,demonstrating the superior performance of GraphICF over several state-of-the-art ICF models and graph-based CF methods.