<p>Recommender systems play a crucial role in addressing information overload, yet traditional centralized approaches raise significant privacy concerns. Federated learning has emerged as a promising solution for privacy-preserving recommendations, but existing methods face challenges such as data sparsity (limited local interactions per client) and data heterogeneity (non-IID user preferences and skewed distributions across clients), which hinder the learning of reliable representations and limit personalization. To address these issues, we propose Federated Recommendations with Personalized Contrastive Learning on Graph Neural Networks (FRecCL), a novel framework that integrates graph neural networks (GNNs) with federated learning. On the client side, we introduce a structural contrastive learning technique that enhances node representations by defining structural neighbors and treating them as positive pairs, thereby mitigating data sparsity. On the server side, we group users into clusters based on their learned representations and compute both cluster-level federated models and a global federated model. Each user then learns a personalized model by combining these two models, effectively addressing data heterogeneity. Extensive experiments on five real-world datasets demonstrate that FRecCL consistently outperforms state-of-the-art methods, achieving an average improvement of 4.44% in HR@10 and 4.76% in NDCG@10 compared to the best baseline. Notably, FRecCL achieves up to 7.92% improvement in HR@10 on the FilmTrust dataset and 7.19% improvement in NDCG@10 on the Amazon-Electronic dataset, showcasing its ability to handle sparsity, heterogeneity, and privacy concerns effectively. Our work advances the field of federated recommendations by providing a robust and scalable solution for personalized, privacy-preserving recommendation systems.</p>