Abstract: With the rapid development of social networks, users hope to obtain more accurate and personalized services. In general, POI recommendation often uses the historical behaviors of a user to recommend the top N POIs and rarely consider the current state of the user. Unlike POI recommendation, successive POI recommendation is more sensitive to user preferences and changes in time and space. In order to alleviate the data sparsity, we make full use of the interaction of time, space and user interest preferences, and propose a successive POI recommendation model called UTeSp. The UTeSp model uses the collaborative influence of multiple interactions to build a model, which can well adapt to the needs of users at different times and different locations. And it can change dynamically. Furthermore, we associate the user's inherent interest preference with the user's friend's influence on the target user, and propose a user-level interest preference based on attention mechanism, which can obtain more accurate user preference results. In addition, a novel TDP_HC algorithm is designed to segment time dynamically. Based on the partial order relationship, we propose two interpretable methods to enhance the learning ability of the model. The two methods can be used in other similar successive POI recommendation models. Experimental results show that the F1-score of UTeSp model on the two real datasets is better than that of several mainstream successive POI recommendation models, and the two partial order methods also show the effectiveness of our model.
Authors: Nan Wang (Heilongjiang University, China); Yong Liu and Peiyao Han (Heilongjiang University，Harbin, China); Xiaokun Li (Postdoctoral Program of Heilongjiang Hengxun Technology Co., Ltd., China); Jinbao Li (Heilongjiang University，Harbin, China)
Email: email@example.com, firstname.lastname@example.org, email@example.com, firstname.lastname@example.org, email@example.com