A new item similarity based on α-divergence for collaborative filtering in sparse data. (15th March 2021)

• Record Type: Journal Article
• Authors: Wang, Yong; Wang, Pengyu; Liu, Zhuo; Zhang, Leo Yu
• Is Part Of: Expert systems with applications. Volume 166(2021)
• Publication Date: 2021-03-15
• Subjects: Similarity measure -- α-divergence -- Collaborative filtering -- Recommendation -- Management information system; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2020.114074
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: The α -divergence is discretized to determine item similarity. The inherent asymmetry of α -divergence is corrected by designing a new value for α . User factor is considered in the design of item similarity. The results indicate that the α -CF improves recommendation quality and efficiency. Abstract: In big data era, collaborative filtering as one of the most popular recommendation techniques plays an important role to promote the development of online trade. Similarity measurement is a core step in collaborative filtering as it not only determines the selection of neighbors but also has a decisive influence on the recommendation quality. However, most of existing similarity measures depend on the co-rated cases(i.e., cases where different users rated the same items or different items were rated by the same users), which usually leads to low data utilization and even poor recommendation results in a sparse dataset. To alleviate this problem, we proposed a new item similarity measure based on α -divergence, which does the computation according to the probability density distribution of ratings and greatly reduces the dependence on co-rated cases. Furthermore, the presented item similarity measure also considers the impact of the absolute number of ratings and the proportion of co-rated cases on the computation results, which effectively improves the accuracy of recommendation. Experiments on three open datasets suggest that the proposed scheme has high prediction accuracy and good adaptability to sparse data. Therefore, it has high potential to be applied in recommender systems.