Mix geographical information into local collaborative ranking for POI recommendation

Point-of-Interest recommendation is a task of personalized ranking prediction on a set of locations for users. The problem of data sparsity is much severe in POI recommendation, as users usually visit only a few POIs concentrated on a limited number of types, relative to the enormous whole POIs. However, he/she has different personalized favors on each type. Based on this phenomenon, we assume the user-POI matrix is locally low-rank instead of globally low-rank, then we put forward to utilize local collaborative ranking (LCR) for POI recommendation, which could mitigate the sparsity of check-in data. Especially, POIs visited by a user always scatter on limited spatial areas, and POI is usually popular in a local scope. There exists spatial local property in users’ check-in behavior. Moreover, to represent the spatial local property, we propose spatial similarity in the first time. With spatial similarity, LCR can find more latent neighborhoods in its local matrices and construct the local matrices much accurately. Besides, user’s preference to POI includes not only general favor but also spatial favor. So spatial favor is introduced in our model. We utilize spatial similarity and spatial favor to mix geographical information into local collaborative ranking seamlessly, proposing our model MG-LCR (Mix Geographical information into Local Collaborative Ranking). Experiments show that, MG-LCR model can reflect users’ preference to POIs more accurately and outperform the state-of-the-art methods.