Tensor Factorization Based POI Category Inference

Abstract

Trajectory data is an important kind of data with different aspects of the user information like demographics, user behavior and activities. Therefore, it is significant and essential to infer point-of-interests (POI) categories from trajectory data for user modeling and user preferences mining in many location-based services (LBS). Recent researches focus more on recommendation and prediction of next POI, which are based on the check-in data. Check-in data is only a partial aspect of the user’s behavior which collected by a certain LBS, while trajectory data describes the user from all around, which can help modeling user’s interest preferences in a great degree. However, due to a deviation between the GPS-coordinate and the actually visited location, it is significant to infer the ultimate POI categories people accessed from trajectory data instead of mapping location coordinates to POIs directly. In this paper, we propose a collaborative inferring framework to analyze the actually visited POI categories from users’ historical trajectory data. Through modeling relationships among the user, time and POI category, the tensor decomposition method can effectively complement the missing data and provides accurate predictions when user trajectory data is absent. Extensive experiments have been conducted with various state-of-the-art baseline on real-world trajectory data, and experiment results have demonstrated the promising performance in this framework.