Bandit-Enhanced Graph Collaborative Filtering via Graph Reconstruction

The recommendation system’s overall goal is to provide personalized suggestions to users rather than recommend popular items. However, standard training paradigms assume that users have static preferences between users and items based on either implicit (click) or explicit (rating) feedback, favoring the model toward popular items. These lead to the effect where popular items are recommended more often and become more popular. The current work has dramatically improved the quality of recommendation by using a graph neural network. However, less attention has been paid to unbalanced recommendation results of hot and cold items. This paper explores the problem of cold and hot items recommendation in a graph-based recommendation system from a new perspective. We propose that the essence of the traditional graph convolution processing is to expand the graph into the tree structure, which loses the original interaction graph’s topological structure information. This loss of information leads to the lack of modeling of users’ preferences for unpopular items; therefore, the recommendation results favor popular items. For this reason, we first introduce Bandit algorithm into the existing graph neural network to capture the graph location information, extract and model the potential purchase tendency of users. It is worth noting that our model is not dependent on a particular graph training method, so it is easy to migrate to all graph-based recommendation systems. We verify the effectiveness of the method on three real datasets, exhibiting substantial improvements (about 8.0% relative improvement on average over the baseline models).