Personalized recommendation in large-scale social platforms faces challenges of privacy preservation, trustworthiness of user contributions, and high carbon footprint due to computationally intensive training. Traditional centralized and federated approaches often overlook user reliability, non-IID (Non-Independent and Identically Distributed) interaction patterns, and environmental sustainability. To address these issues, this paper proposes TA-FURL (Trust-Aware Federated User Representation Learning) framework that integrates probabilistic trust inference, carbon-aware federated optimization, and self-supervised contrastive learning for robust and eco-efficient recommendations. Local embeddings are trained using variational autoencoders, and encrypted updates are aggregated with trust-weighted averaging under strict carbon budgets, ensuring privacy, reliability, and energy efficiency. Extensive experiments on MovieLens-1 M, Epinions, and CiaoDVD datasets demonstrate that TA-FURL outperforms existing federated baselines, achieving up to 6.8% improvement in NDCG@10 and 5.9% in Recall@20, while reducing communication cost by 21% and estimated carbon emissions by 11.9%. Ablation studies confirm the critical contributions of trust regularization, carbon-aware optimization, and contrastive learning. Scalability tests with up to 200 clients and cross-dataset transfer experiments further validate the framework’s generalizability. Overall, TA-FURL provides a practical, privacy-preserving, trust-aware, and carbon-efficient solution for next-generation personalized recommendation in heterogeneous social platforms.
Rawat et al. (Mon,) studied this question.