Reduce unrelated Knowledge through Attribute Collaborative signal for knowledge graph recommendation. (1st September 2022)

• Record Type: Journal Article
• Authors: Qian, Fulan; Zhu, Yuhui; Chen, Hai; Chen, Jie; Zhao, Shu; Zhang, Yanping
• Is Part Of: Expert systems with applications. Volume 201(2022)
• Publication Date: 2022-09-01
• Subjects: Recommender system -- Knowledge graph -- Collaborative signal -- Attention mechanism; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.117078
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: K nowledge graph (KG), as an auxiliary information, plays an important role in the recommendation system, which effectively solves the sparsity and cold start problems of collaborative filtering algorithms. The recommendation algorithm that introduces the propagation mechanism on the KG has been a great success, it enriches the representation of users and items by aggregating multi-hop neighbors. However, the existing KG-based propagation recommendation algorithm aggregating all entity information cannot guarantee the improvement of recommendation results, because entity information in KG is not all helpful to recommend appropriate items to users. Indiscriminately aggregating the entity information in the neighborhood allows the learned embedding representation to be influenced by its unrelated entities. In this paper, we propose a new model named R educe unrelated K nowledge through A ttribute C ollaborative signal (RKAC). Compared to other KG-based propagation methods, RKAC offers a new concept of combining item attributes with collaborative signals to reduce information about unrelated entities. Specifically, the initial entity set of users was obtained by collaborative signals and the initial entity set of items was obtained by filtering redundant collaborative signals based on item attributes, and then they were propagated on KG as seeds to acquire multi-hop neighbor entities. Finally, domain entities of different importance were gathered through attention mechanism to obtain more accurate embedding representation of entities. Experimental results on four benchmark datasets of music, book, movie and restaurant show that the AUC of RKAC on CTR prediction increases by 1.4%, 1.3%, 0.8% and 0.5% respectively, compared with the state-of-the-art existing approaches. Highlights: A reduce unrelated knowledge through attribute collaborative recommender method. Reduce redundant knowledge information in knowledge graph through attribute. Aggregate the neighborhood information with multi-hop and high correlation. Using attention mechanism to distinguish the importance of nodes.