Multi-perspective Hierarchical Dirichlet Process for Geographical Topic Modeling

The pervasion of location acquisition technology has strongly propelled the popularity of geo-tagged user-generated content (UGC), which also raises new computational possibility for investigating geographical topics and users’ spatial behaviors. This paper proposes a novel method for geographical topic modeling by combining text content with user information and spatial knowledge. Topics are estimated as the interests of users and features of locations. The joint modeling of the three heterogeneous sources (1) leads to high accuracy in predicting visit behaviors driven by personal interests, (2) discovers coherent topic representations for topic modeling, (3) enables the recommender system to suggest interpretable locations. Our framework is flexible to incorporate new dimensions of data such as temporal information without substantially changing the model structure. We also experimentally demonstrate the limitations of the traditional assumption that a topic is selected considerably dependent on the location. In many cases, the published topics are mainly affected by the user’s interests rather than the current location. Our model discriminates these two scenarios. Through employing hierarchical Dirichlet process, we also need not predefine the number of topics like other mixture models. Experiments on three different datasets show that our model is effective in discovering spatial topics and significantly outperforms the state of the art.