Spatio-Temporal Aware Next Point-of-Interest Recommendation with Privacy Preserving

With the increasing popularity of mobile Internet, recommending the next Point-of-Interest (POI) to users based on their current spatio-temporal locations has become an essential personalized service in Location-Based Social Networks (LBSNs). However, this convenience comes at the cost of privacy risks. To address this issue, federated learning has emerged as a popular privacy-preserving solution. Nevertheless, it faces challenges such as high communication complexity, vulnerability to gradients leakage attacks, and Non-IID data. In view of this challenge, we propose a privacy-preserving next POI recommendation framework in this paper, based on federated learning in MEC environment. Concretely, we first design a lightweight next POI recommendation algorithm based on spatio-temporal aware self-attention network (STS-POIRec), which combines the geographical distances, time intervals, and relative positions of users’ check-in POIs with self-attention network to learn users’ dynamic spatio-temporal preferences. Then, we present a federated learning approach (MEC-Fed) to train the proposed STS-POIRec in MEC environment. Using three real-world check-in datasets, the experimental results show that the proposed STS-POIRec achieves good performance with lower computing complexity and communication costs. The MEC-Fed approach can significantly alleviate the problem of Non-IID data, and obtain performance close to that of centralized learning in next POI recommendation.