A graph convolutional network to improve item recommendation by incorporating bundle-based side-information with multi-level propagations

Side-information, such as bundle, type, or brand, can be used to improve item recommendations. Use of graph convolutional networks (GCN) to calculate embeddings of side-information through user-side-item heterogeneous networks is common in the recommendation domain. However, current GCN-based methods largely ignore the limitations of bundle side-information. This is for two reasons: in some bundles, interaction with users or items is sparse; while in others, contributions of items cannot be estimated accurately due to irrelevant and noisy interactions. To overcome these limitations, we propose Graph Convolutional Network incorporating Bundle-based Side-Information (GCN-BSI). Unlike earlier studies, which model user, item and side-information into a unified graph, this model reduces the negative influence of bundle side-information by splitting the graph into three-level (lower, middle and upper) propagation models and incorporating these models into a unified framework by adopting different propagation strategies at different levels. This framework can make better use of bundle semantic information by iteratively optimising models from lower to upper levels, thereby controlling the quality of propagated information. This refined approach can further improve the performance of item recommendations. In a series of experiments, GCN-BSI was compared with eight state-of-the-art baselines using data from NetEase and SteamGame. GCN-BSI showed a significant improvement. An ablation test and case studies further indicated that the optimised solution was better at capturing user–item correlations from specific side-information. The code and data can be visited at: https://github.com/zhongshsh/GCN-BSI