Joint Topic-Semantic-aware Social Matrix Factorization for online voting recommendation

• Published in: Knowledge-based systems Vol. 210; p. 106433
• Main Authors: Wang, Jia, Wang, Hongwei, Zhao, Miao, Cao, Jiannong, Li, Zhuo, Guo, Minyi
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 27.12.2020; Elsevier Science Ltd
• Subjects: Factorization; Matrix factorization; Online voting; Recall; Recommender systems; Semantics; Topic-enhanced word embedding; Voting; Words (language); Matrix factorization; Topic-enhanced word embedding; Recommender systems; Online voting
• ISSN: 0950-7051; 1872-7409
• DOI: 10.1016/j.knosys.2020.106433

Social voting is an emerging new feature in online social platforms, through which users can express their attitudes and opinions towards various interested subjects. Since both social relations and textual content decide the votes propagation, the diverse sources present opportunities and challenges for recommender systems. In this paper, we jointly consider these two factors for the online voting recommendation. First, we conduct feature learning on the vote content. Note that the vote questions are usually short and contain informal expressions, existing text mining methods cannot handle it well. We propose a novel topic-enhanced word embedding (TEWE) method, which learns the word vectors by considering both token-level semantics and document-level mixture topics. Second, we propose two Joint Topic-Semantic-aware Social Matrix Factorization (JTS-MF) models, which fuse social relations and textual content for the vote recommendation. Specifically, JTS-MF1 directly identifies the interaction strength to calculate the similarity among users and votes, while JTS-MF2 aims to preserve inter-user and inter-vote similarities during matrix factorization. Extensive experimental results on real online voting dataset show the effectiveness of our approaches against several state-of-the-art baselines. JTS-MF1 and JTS-MF2 models surpass the matrix factorization based method, with 25.4% and 57.1% improvements in the top-1 recall, and 59.12% and 25.1% improvements in the top-10 recall.