Using trajectories for collaborative filtering-based POI recommendation

Current mobile guides often suffer from the following problems: a long knowledge acquisition process of recommending relevant Points of Interest (POIs), the lack of social navigation support, and the challenge of making implicit user-generated content (e.g., trajectories) useful. Collaborative filtering (CF) is a promising solution for these problems. This article employs CF to mine GPS trajectories for providing Amazon-like POI recommendations. Three CF methods are designed: simple_CF, freq_CF (considering visit frequencies of POIs), and freq_seq_CF (considering both user’s preferences and spatio-temporal behaviour). With these, services like “after visiting ..., people similar to you often went to ...” can be provided. The methods are evaluated with two GPS datasets. The results show that the CF methods can provide more accurate predictions than simple location-based methods. Also considering visit frequencies (popularity) of POIs and spatio-temporal motion behaviour (mainly the ways in which POIs are visited) in CF can improve the predictive performance.