Heterogeneous information network-based interest composition with graph neural network for recommendation

Heterogeneous information networks (HINs) are widely applied to recommendation systems due their capability of modeling various auxiliary with meta-paths. However, existing HIN-based models usually fuse the from meta-paths by simple weighted sum or concatenation, which limits performance improvement because it lacks interest compositions among In this article, we propose an Interest Composition model for Recommendation (HicRec). Specifically, user and item representations learned a graph neural network on both structure features in each meta-path, parameter sharing mechanism is utilized here ensure that same latent space. Then, users' interests pair related calculated combination representations. The composed obtained single intra- inter-meta-paths recommendation. Extensive experiments conducted three real-world datasets results demonstrate our proposed HicRec outperforms baselines.