Attention-aware metapath-based network embedding for HIN based recommendation. (15th July 2021)

• Record Type: Journal Article
• Authors: Yan, Surong; Wang, Haosen; Li, Yixiao; Zheng, Yuan; Han, Long
• Is Part Of: Expert systems with applications. Volume 174(2021)
• Publication Date: 2021-07-15
• Subjects: Heterogeneous information network -- Recommender system -- Network embedding -- Deep learning -- Attention mechanism; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2021.114601
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: Heterogenous information network based recommendation is investigated. An attention-aware metapath-based network embedding approach is proposed. Each metapath is modeled as a weighted homogenous information network. A self-attention mechanism generates integrated representations of users and items. Deep neural network methods are used in the final stage of prediction. Abstract: Heterogeneous information network (HIN) attracts increasing attention from the communities of recommender systems. HIN based recommendation methods can help overcome the difficulties of data sparsity and cold start. The majority of the existing HIN based recommendation methods use path-based semantic similarity between users and/or between items on HINs. However, the existing HIN based recommendation methods using metapath disregard the semantic differences among multiple metapaths (i.e., inter-metapaths) and the influence differences among neighbor pairs in each individual metapath (i.e., intra-metapaths). To solve these problems, we propose an attention-aware metapath-based network embedding for HIN based recommendation. To obtain additional semantic information, our method generates multiple metapath-based weighted homogeneous networks to model the auxiliary information of users and items of HIN. Thereafter, we design a novel self-attention integration to integrate multiple semantic information from multiple weighted homogenous information networks. Lastly, we utilize three deep neural network methods to model the implicit relations between users and items for the rating prediction task. Experimental results of three real-world datasets demonstrate that the proposed model outperforms existing state-of-the-art recommendation methods, solves the data sparsity problem, and models the multiple semantic information of users and items.