Context-aware Location Recommendations with Tensor Factorization

Location-based social networks (LBSNs) enable users not only to record where and when they go using check-ins but also to publish their ratings and comments of these locations. All information can be utilized to infer users' interests and to generate personalized location recommendations for different individuals. In this process, temporal context plays a significant role, because life patterns of users vary greatly and different locations have different proper visiting time. Therefore, in this paper, we propose a top-k location recommendation scheme, considering the influence of temporal context. To be specific, we first model user-location-time relations with a three dimensional tensor. Then, we extract a user-user similarity matrix and a location feature matrix to relieve data sparsity problem and to increase the prediction accuracy of estimated ratings during the tensor decomposition process. Through a context-aware tensor decomposition approach, we can obtain any user's rating for any location in any time slot. Moreover, we partition a large area into several small ones and adopt the Threshold Algorithm (TA) to improve efficiency of online recommendations. We evaluate our method in a real-world dataset and the results demonstrate the effectiveness of our method.