CoCo-POI: Collaborative Contrastive Learning for Point-of-Interest Recommendation

Next Point-Of-Interest (POI) recommendation is essential for optimizing user travel experiences and enhancing platform revenues by providing users with potentially attractive location options. Traditional centralized systems depend on extensive datasets of prior POI check-ins, raising significant privacy and computational concerns. Centralized and decentralized Federated Learning (FL) systems face key challenges, including balancing personalization and privacy, reducing communication overhead, optimizing computational costs, and addressing data sparsity. To tackle these challenges, we propose the CoCo-POI algorithm, a collaborative contrastive learning framework for POI recommendation that leverages a Federated Knowledge Distillation (FKD) approach. CoCo-POI utilizes contrastive learning for data augmentation on local devices, and employs a neighbor knowledge aggregation method based on similar soft decisions. Additionally, it uses sparse compression techniques to decrease the size of local knowledge sets, thereby communication costs. Experimental results demonstrate that CoCo-POI achieves superior outcomes in the next POI recommendation tasks while maintaining efficiency in resource utilization.