A precision-centric approach to overcoming data imbalance and non-IIDness in federated learning

Federated learning (FL) enables decentralized model training, but the distribution of data across devices presents significant challenges to global model convergence. Existing approaches risk losing the representativeness of local models after model aggregation, calling for a more efficient and robust solution. In this study, we address the model aggregation challenge within FL by focusing on elevating the performance of the global model amidst class imbalance and non-independent, non-identically distributed data. We aim to train a global model collaboratively that represents all participating nodes, promoting fairness and ensuring adequate representation of all classes in the model. We propose redistributing local model weights based on their precision-based contributions to each class to enhance the performance and communication efficiency of federated thermal comfort prediction. Our proposed method can assist in allocating more resources and attention to nodes with high precision for underrepresented classes, thereby improving the global model overall performance and fairness. Furthermore, our framework leverages the virtualization capability of digital-twin to enable the dynamic registration and participation of nodes in the federated learning process in real-time. The developed digital-twin framework allows for real-time monitoring and control of the decentralized training. Through our evaluation on a real dataset, we showcase noteworthy enhancements in accuracy and communication efficiency when compared to existing methods. Our evaluation shows that the proposed Class Precision-Weighted Aggregation technique (Fed-CPWA) outperforms Federated Averaging, with higher accuracy of 82.85% and lower communication costs by 25%. Our contribution represents a significant stride towards sustainable thermal comfort modeling, further advancing the development of equitable and resilient federated learning techniques.