Author(s)

Jin Hu, Jinghua Zhu and Heran Xi

Corresponding Author

Heran Xi

ABSTRACT

As considerable amounts of POI check-in data have been accumulated,successive point-of-interest (POI) recommendation is increasingly popular. Due to the scarcity of check-in data, how to accurately capture user preferences and poor interpretation in the temporal model are two major problems faced by traditional recommender systems.To this end, we propose a new category-aware social-geographic deep model.Our model consists of a pre-training module, an encoder module and a filter module,designed to learn the POI embedded representation incorporating social relations and geographical influences, and serves as the initial values for two LSTM based encoders.The two encoders are designed to capture user preferences in the POI category and user preferences for the POI at the current moment.The filter module is used to filter the candidate POIs through the category of POIs.Finally, we sort the candidate set by considering three specific dependencies: user-poi, user-poi category, and current preferences of POI users.Experiments were conducted on two large real datasets.Experimental results show that our CATDM is superior to existing models.

KEYWORDS

lbsn, poi, recommendation, category-aware