Joint Geo-Spatial Preference and Pairwise Ranking for Point-of-Interest Recommendation

Recommending users with preferred point-of-interests (POIs) has become an important task for location-based social networks, which facilitates users' urban exploration by helping them filter out unattractive locations. Although the influence of geographical neighborhood has been studied in the rating prediction task (i.e. regression), few work have exploited it to develop a ranking-oriented objective function to improve top-N item recommendations. To solve this task, we conduct a manual inspection on real-world datasets, and find that each individual's traits are likely to cluster around multiple centers. Hence, we propose a co-pairwise ranking model based on the assumption that users prefer to assign higher ranks to the POIs near previously rated ones. The proposed method can learn preference ordering from non-observed rating pairs, and thus can alleviate the sparsity problem of matrix factorization. Evaluation on two publicly available datasets shows that our method performs significantly better than state-of-the-art techniques for the top-N item recommendation task.