Side Information Fusion for Recommender Systems over Heterogeneous Information Network

Collaborative filtering (CF) has been one of the most important and popular recommendation methods, which aims at predicting users’ preferences (ratings) based on their past behaviors. Recently, various types side information beyond explicit ratings users give to items, such as social connections among metadata have introduced into CF shown be useful for improving performance. However, previous works process different separately, thus failing capture correlations that might exist across them. To address this problem, in work, we study application heterogeneous network (HIN), offers a unifying flexible representation information, enhance CF-based methods. face challenging issues HIN-based recommendation, i.e., how similarities complex semantics between items HIN, effectively fuse these improve final issues, apply metagraph similarity computation solve fusion problem with “matrix factorization (MF) + machine (FM)” framework. For MF part, obtain user-item matrix from each then low-rank approximation latent features both items. FM Group lasso (FMG) obtained part train recommending model and, same time, identify metagraphs. Besides FMG, two-stage method, further propose an end-to-end hierarchical attention fusing, metagraph-based recommendation. Experimental results four large real-world datasets show two proposed frameworks significantly outperform existing state-of-the-art methods terms