A Spatial User Similarity Measure for Geographic Recommender Systems

Recommender systems solve an information filtering task. They suggest data objects that seem likely to be relevant to the user based upon previous choices that this user has made. A geographic recommender system recommends items from a library of georeferenced objects such as photographs of touristic sites. A widely-used approach to recommending consists in suggesting the most popular items within the user community. However, these approaches are not able to handle individual differences between users. We ask how to identify less popular geographic objects that are nevertheless of interest to a specific user. Our approach is based on user-based collaborative filtering in conjunction with an prototypical model of geographic places (heatmaps). We discuss four different measures of similarity between users that take into account the spatial semantic derived from the spatial behavior of a user community. We illustrate the method with a real-world use case: recommendations of georeferenced photographs from the public website Panoramio. The evaluation shows that our approach achieves a better recall and precision for the first ten items than recommendations based on the most popular geographic items.