UC San Diego

With the emerging of various location-based social networks (LBSNs), the study on user mobility patterns and many related tasks have become heated research topics, such as personalized location recommendation and friend recommendation. Many factors affect users’ behavior patterns, such as geographical influence, temporal effect and semantic effect. However, most of the previous work on modeling user trajectories lacks consideration on treating these factors from a graph perspective, therefore fails to capture the potential correlations among the rich context. In this thesis, we demonstrate that using a heterogeneous graph-based model to jointly embed user and POI attribute networks with a unified framework can well preserve the network properties. Multiple factors are embedded into a shared low-dimensional latent space where their joint effect and potential correlations can be well captured. We conduct extensive experiments on large real-world datasets to evaluate our model performance on several major tasks in LBSNs. The experimental results highlight the versatility of our method which shows higher recommendation effectiveness compared with the state-of-the-art baselines. In addition, an online updating strategy is proposed to incorporate new visiting records and dynamically track users’ latest preference in linear time. We also show that this framework has the inherent ability to handle cold-start recommendations, which is a non-trivial task considering the network sparsity of LBSNs. The scalability and flexibility of our framework indicate that this method is promising to be put into practical use.