The Next Point-Of-Interest (POI) recommendation has gained prominence with the rapid expansion of Location-Based Social Networks (LBSNs), due to its effectiveness in leveraging historical check-in data to predict users' next visiting POIs. While current centralized POI recommenders based on Recurrent Neural Networks (RNNs) and Graph Neural Networks (GNNs) have achieved notable performance, these approaches require users to upload sensitive check-in data, raising privacy concerns. Additionally, high-latency responses, dependent on network quality and server stability, further limit their practicality. Federated learning has emerged as a privacy-preserving alternative by aggregating locally trained models into a centralized federated model, which is then distributed for inference. However, this approach still involves uploading user-trained models, risking data exposure, disadvantaging minority groups, and neglecting diverse regional interests. To address these issues, decentralized on-device POI recommender systems have been proposed, enabling local training and inference while enhancing models through direct knowledge exchange within groups. Despite its promise, this framework faces challenges such as data sparsity, model heterogeneity, cold-start problems, and advanced privacy threats. This thesis introduces several innovations to tackle these challenges: (1) Decentralized Collaborative Learning Framework for POI Recommendations (DCLR): Users train personalized models locally, and those models are further enhanced through knowledge exchange with neighbors identified by categorical similarity and geographic proximity. (2) Diffusion-Based Cloud-Edge-Device Collaborative Learning Framework for POI Recommendations (DCPR): A well-trained model is deployed on devices for zero-shot recommendations or further fine-tuning. (3) Model-Agnostic Collaborative Learning Framework for POI Recommendations (MAC): A novel knowledge distillation approach facilitates collaboration between heterogeneous models, allowing users to tailor configurations and focus on context-relevant POIs. (4) Physical Trajectory Inference Attack and Defense in Decentralized POI Recommendations (PTIA): To explore privacy concerns in decentralized POI recommendations, we propose a physical trajectory inference attack (PTIA), aimed at identifying users' interacted POIs. To counter PTIA, we design an adversarial game (AGD) to eliminate sensitive POIs and their implicit associations in shared models, mitigating privacy risks in collaborative learning. These contributions advance decentralized POI recommendations by improving accuracy, user personalization, and privacy, paving the way for more secure and efficient systems.