Context-aware next point-of-interest recommendation with uncertain check-ins

The rapid development of next point-of-interest (POI) recommendation benefits from a large number of check-ins shared by users in location-based social networks (LBSNs), such as Foursquare and Yelp, which helps users explore their surroundings. Accordingly, most existing studies assume that such check-ins reflect users' exact visits, i.e., certain check-ins. In reality, due to the privacy concern or the bias of the indoor navigation of GPS, users may leave some uncertain check-ins at a collective POI, e.g., a shopping mall containing multiple individual POIs. Such uncertain check-ins bring the challenges of learning user preference and complete transition patterns, which are rarely investigated by the existing next POI recommendation studies. Therefore, in this dissertation, we focus on a new research problem that aims to recommend next individual POIs with uncertain check-ins, and develop a series of effective algorithms by exploiting context information mined from users' historical check-in behaviors to resolve this problem. In the next POI recommendation scenario, exploiting users' preference transitions over the categories of POIs, i.e., category transition, is of significance in improving recommendation accuracy. Moreover, POIs are typically organized by a category hierarchy (CH). As a result, CH can better describe the correlations between POIs (categories) and categories, which shows potential in easing the cold start issue in the recommendation task. Existing studies either exploit the category transition or CH to improve the recommendation performance. However, they lack consideration of both effects in a unified manner. In order to take advantage of their joint effects for better recommendation under the uncertain check-in scenario, we first propose a HCT framework, which exploits hierarchical category transitions for better multi-granularity user preference transition learning. In this way, HCT predicts users' preferred categories inside collective POIs. As bounded to specific categories, HCT further adopts hierarchical dependencies in CH to capture the semantic relatedness of POIs, thus easing the cold start issue. Empirical studies show the superiority of HCT against state-of-the-art algorithms. Despite the pioneer effort HCT can address the challenge of uncertain check-ins to some extent, it fails to well characterize users' underlying activities over uncertain check-ins and model the interplay between sequential activities and locations by considering the spatiotemporal context. Therefore, we devise an interactive multi-task learning framework – iMTL, which introduces: a temporal-aware activity encoder equipped with fuzzy characterization to unveil the latent activity transition patterns; a spatial-aware location preference encoder to capture the latent location transition patterns; and a task-specific decoder to adopt the learned latent transition patterns and enhance both activity and location prediction tasks in an interactive manner. Extensive experiments on real-world datasets demonstrate that iMTL consistently outperforms the state of the arts. Although the proposed HCT and iMTL have shown the capability of easing the issue of uncertain check-ins, they suffer from the inability of recommending accurate POIs when they fail to predict user’s accurate activity (i.e., immediate preference). In fact, most existing next POI recommenders (i.e., static methods) face the inherent limitations of capturing users’ accurate immediate preferences and adapting to their feedback regarding the recommendation results. The development of conversational recommender system (CRS) techniques brings great potential in resolving the limitations of such static recommenders. We further propose a conversation-based adaptive relational translation (CART) approach for next POI recommendation over uncertain check-ins. It is equipped with recommender and conversation modules to interactively acquire a user's immediate preference and make dynamic recommendations. Specifically, the recommender built upon the adaptive relational translation method performs location prediction via modeling both sequential behaviors and the immediate preference received from conversations; and the conversation module aims to achieve successful recommendations in fewer conversation turns by learning a conversational strategy, whereby the recommender can be updated via the user's feedback. Experiments on real-world datasets show the superiority of our proposed CART over state-of-the-art algorithms. To sum up, in this dissertation, we propose a series of recommendation approaches by exploiting context information mined from users' check-in behaviors for more accurate next POI recommendations under the uncertain check-in scenario.Doctor of Philosoph