Learning from contextual information of geo-tagged web photos to rank personalized tourism attractions

This paper proposed a method that fully exploits contextual information of geo-tagged web photos to recommend tourism attractions to a user according to his personal interest and current time and location. The proposed method first detects tourism attractions from geo-tags, and estimates their popularity with users' photo quantity. Photos' taken time is used to discover temporal fluctuation of attractions' popularity and distance of consecutive photos is exploited to model the spatial influence to user's travel behavior. Photos' textual and visual information are used to reveal users' personal interests. Collaborative filtering is also adopted in the recommendation process. With all these contextual information, our method predicts a user's preference to a certain attraction from different aspects, and automatically combines the prediction scores to give the final recommendation result with a learning to rank model. Experiments on Panoramio dataset show that our method performs better than the state-of-the-art method, especially for users with little traveling history.