Recommendation in Location-based Social Networks

Recommender systems have become popular tools to select relevant personalized information for users. With the rapid growth of mobile network users, the way users consume Web 2.0 is changing substantially. Mobile networks enable users to post personal status on online social media services from anywhere and at anytime. However, as the volume of user activities is growing rapidly, it is getting impossible that for users to read all posts or blogs to catch up with the trends. Similarly, it is hard for producers and manufactures to monitor consumers and figure out their tastes. These needs inspired the emergence of a new line of research, recommendation in location-based social networks, i.e., building recommender systems to discover and predict the behavior of users and their engagement with location-based social networks. Extracted users' interests and their spatio-temporal patterns clearly provide more detailed information for producers to make decisions to supply their consumers. In this thesis, we address the problem of recommendation in location-based social networks and seek novel methods to improve limitations of existing techniques. We first propose a spatial topic model for top-k POI recommendation problem, and the proposed model discovers users' topic and geographical distributions from user check-ins with posts and location coordinates. Then we focus on mining spatio-temporal patterns of user check-ins and propose a spatio-temporal topic model to identify temporal activity patterns of different topics and POIs. In our next work, we argue that all existing social network-based POI recommendation models cannot capture the nature of location-based social network. Hence, we propose a social topic model to effectively exploit a location-based social network. Finally, we address the problem of determining the optimal location for a new store by considering it as a recommendation problem, i.e., recommending locations to a new store. Latent factor models are proposed and proved to perform better than existing state-of-the-art methods.