GIST: A generative model with individual and subgroup-based topics for group recommendation. (15th March 2018)

• Record Type: Journal Article
• Authors: Ji, Ke; Chen, Zhenxiang; Sun, Runyuan; Ma, Kun; Yuan, Zhongjie; Xu, Guandong
• Is Part Of: Expert systems with applications. Volume 94(2018)
• Page Start: 81
• Page End: 93
• Publication Date: 2018-03-15
• Subjects: Group recommendation -- Group activity -- Decision making -- Topic model -- Recommender systems; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2017.10.037
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: A novel Topic-based model is proposed to make group recommendations. A new type of subgroup topic is introduced to enrich group activity analysis. A new solution to choice aggregation is designed to inferring group decision. The link information of group members is used to optimize the weights. Results on real-life data are presented to illustrate the performance of our model. Abstract: In this paper, a Topic-based probabilistic model named GIST is proposed to infer group activities, and make group recommendations. Compared with existing individual-based aggregation methods, it not only considers individual members' interest, but also consider some subgroups' interest. Intuition might seem that when a group of users want to take part in an activity, not every group member is decisive, instead, more likely the subgroups of members having close relationships lead to the final activity decision. That motivates our study on jointly considering individual members' choices and subgroups' choices for group recommendations. Based on this, our model uses two kinds of unshared topics to model individual members' interest and subgroups' interest separately, and then make final recommendations according to the choices from the two aspects with a weight-based scheme. Moreover, the link information in the graph topology of the groups can be used to optimize the weights of our model. The experimental results on real-life data show that the recommendation accuracy is significantly improved by GIST comparing with the state-of-the-art methods.