Discovery of User Preferences from Big Geospatial Data Using Topic Models

It is becoming common to apply location information for actual problems such as automatic routing, creating tour, and the prediction of user activity. In many of these cases, they use a geodata that is based on a user's self check-in. In this study, to focus on a user's daily behavior deeply, we use a large volume of geospatial data, which is detected regularly and automatically from a cell phone's GPS system and the volume of data is totally different from the self check-in based. We can comprehend users activity in more detail than self check-in based data. The main purpose of this research is to extract and interpret user preferences from big geospatial data. Because the auto-detected data is big and spatial, it is necessary to process and analyze it for transformation to a value. Firstly, we extract the spots where people visited and stayed from raw geodata. Next, we categorize spots using the spot information to reduce the dimension. Finally, we interpret user preferences by applying a topic model to the matrix of users and the categorized spot. Our results show that we can interpret user preferences as topics from big geospatial data using dimension reduction and a topic model. In addition, we verify a performance of methods that use the topics to predict purchase. We found that the models that use topics outperforms others on the real prediction problem.