McHa: a multistage clustering-based hierarchical attention model for knowledge graph-aware recommendation

Knowledge graph-aware recommendation has become an important research topic in recent years. The user preference representation, which preserves the user’s taste towards items (e.g., movies, books.), is obtained through aggregating the information of entities or attributes in knowledge graphs directly. However, the fine-grained heterogeneity information, which can be derived from the groups of items or entities, remains barely exploited in the process of encoding the user interaction intention for the items. To fill up this gap, we propose a Multistage Clustering-based Hierarchical Attention (McHa) model to capture the user preference representation. In our work, we first group the items and their neighboring entities in the knowledge graph into item clusters and entity clusters (jointly referred to as multistage clusters), respectively. Then, the user preference representation is obtained by hierarchically aggregating the heterogeneity information derived from the multistage clusters with the weights generated by the hierarchical attention layers. We conduct extensive experimental comparisons with baselines and the variants. The experimental results indicate that McHa has achieved state-of-the-art performance on three benchmark datasets in two scenarios.