Metapath-guided dual semantic-aware filtering for HIN-based recommendation

Surong Yan, Haosen Wang, Yixiao Li, Chunqi Wu, Long Han, Chenglong Shi, Ruilin Guo

Abstract: Many heterogeneous information network (HIN)-based recommendation methods leverage the semantic and structural features of metapath to improve the recommendation performance. However, the existing HIN-based recommendation methods using metapath still suffer from two challenges: (1) HINs in industrial recommendation scenarios usually have a very large scale and contain much redundant or noisy structural information, which may damage the efficiency and effectiveness of the recommendation model. (2) HINs include rich metapath semantic information that may be noisy and irrelevant to downstream tasks. To address the above two challenges, we propose a metapath-guided dual semantic-aware filtering for HIN-based recommendation from two perspectives: intra-metapath and inter-metapath (called MFGRec). Our model first develops a neighbor filtering method within metapath-guided attribute networks to generate tailored metapath-guided attribute networks for filtering irrelative or noise neighbors of intra-metapath. Moreover, our model designs a semantic-aware filtering-based fusion method using a novel adaptive multi-head sparse attention mechanism to automatically discard the irrelative metapath-guided attribute networks for each user-item interaction pair and assign personalized weight to the selected valuable networks for distinguishing the semantic differences of inter-metapath. In general, MFGRec filters a large amount of noise and irrelevant information from intra-metapath and inter-metapath perspectives, which significantly improves the scalability and accuracy of the recommendation framework. Furthermore, experimental results on three publicly accessible datasets and nine baselines demonstrate that our model achieves higher accuracy of recommendation and lower runtime costs compared with existing state-of-the-art methods.