Graph-based point-of-interest recommendation on location-based social networks

Recent decades have witnessed a high-speed development of urban area with a large amount of POIs (point-of-interests) being built. A POI is a place with some functionalities in a location-based social network (LBSN), such as restaurants and movie theatres. In a LBSN, users can report their geographical locations and experiences explicitly via check-ins. However, the huge amount of heterogeneous information in LBSNs brings tremendous challenges to develop an effective POI recommender system. Due to the extreme heterogeneity, user check-in decision exhibits two critical properties: (i) diversity; and (ii) imbalance. The diversity property is that the choice of visiting a POI is often jointly influenced by multiple factors, such as geographical and social factors. Meanwhile, the imbalance property represents that various influential factors carry different levels of importance for user check-in decisions. To capture both properties, this thesis proposes more advanced POI recommender systems by considering two perspectives: (i) representation: model the heterogeneous information in a unified data structure; and (ii) methodology: develop more effective algorithms to exploit the data structure to facilitate the POI recommendation generation. For the representation, this thesis explores the graph-based techniques from homogeneous graph to heterogeneous graph with the aim to embed various types of data in a unified space; for the methodology, this thesis proposes a series of approaches from random walk based method, to latent factor model, and deep learning model to effectively employ the graph structures for both general and next POI recommendation tasks. The first study proposes a topic-sensitive POI recommendation model with a spatial awareness model (TSLRS) to exploit the geographical and content factors. Specifically, a homogeneous graph consisting of users is built with user topic preferences based on the textual information of POIs. The neighbors are discovered for each user by a topic-sensitive random walk over the graph. The opinions of neighbors are further aggregated to infer the preference score for a POI. Finally, the geographical factor is also used in the neighbor discovery process to find nearby neighbors. However, the homogeneous graph in first study projects the heterogeneous information into a homogeneous representation, leading to information loss. Hence, the second study designs a novel heterogeneous graph (AGS-IG) by fusing various relations among users, POIs and aspects from user reviews. Then, a novel graph-based ranking algorithm (AGS- RW) is proposed based on personalized PageRank (PPR) and meta-paths to model the diversity property by a full exploitation of both the heterogeneous graph structure and the semantic relations of AGS-IG. Despite the exploitation of multiple factors, AGS-RW fails to model the imbalance of user check-in decisions. Thus, the third study develops a matrix factorization framework (AGS-MF) to effectively model both properties. First, AGS-IG is also adopted to represent the heterogeneous information in a common space. Then, an efficient meta-path based random walk over AGS-IG is developed to find relevant neighbors of each user and POI based on multiple factors, which are further incorporated into AGS-MF. By doing so, AGS-MF not only models multiple factors, but also learns the personalized weights for each individual user and POI. Thus, both the diversity and imbalance properties can be captured in a unified manner. For next POI recommendation task, existing studies mainly model the sequential regularity of check-in sequences, but suffer from the severe data sparsity issue, making it extremely difficult to capture the transitional patterns between POIs. Thus, the fourth study proposes an recurrent model (ARNN) to jointly model both the sequential regularity and transition regularities of neighbors. Specifically, a meta-path based random walk over a novel knowledge graph is proposed to discover POI neighbors based on the heterogeneous factors. The transition regularities of various neighbors are integrated via the attention mechanism, which seamlessly cooperates with the sequential regularity as a unified recurrent framework. In summary, a series of recommendation approaches have been proposed for both the general and next POI recommendation tasks, which exploit graph-based techniques to represent the heterogeneous information in a unified space for more effective POI recommendation. The extensive experiments demonstrate the superiority of the proposed approaches over state-of-the-art techniques.Doctor of Philosoph