CA-SparseFL: Context-Aware Sparse Federated Learning for Energy-Efficient Urban Air-Quality Intelligence

Arcaya-Coaquira W.E.; Sarmiento-Calisaya E.; Cunha W.L.; Henry Mamani-Aliaga A.; Bustincio R.W.C.; Bittencourt L.F. · Proceedings - 2026 22nd International Conference on Distributed Computing in Smart Systems and the Internet of Things, DCOSS-IoT 2026 (2026) · DOI: 10.1109/dcoss-iot69657.2026.00175

Resumen

Urban sensing infrastructures produce geographically distributed and privacy-sensitive data streams; however, training models on these data is constrained by communication costs, heterogeneous edge energy, and spatial representativeness. Federated learning (FL) mitigates the need to centralize raw data, but standard client selection methods often favor well-powered or low-cost nodes. This study introduces CA-SparseFL, a context-aware, sparse FL framework designed for urban air-quality intelligence, specifically for PM2.5 forecasting. CA-SparseFL incorporates a lightweight binary integer optimization layer for client selection, top-k sparse update transmission, and a fairness-aware participation term that discourages repeated selection of the same clients. The framework is evaluated using the Beijing Multi-Site Air Quality dataset, modeling 12 monitoring stations as federated clients, each with a 6,000-hour continuous window. The evaluation compares dense random FL, sparse random FL, energy-only scheduling, and CA-SparseFL, and includes ablations of components and a sparsity sensitivity analysis. Results indicate that CA-SparseFL reduces communication by 50% compared with dense transmission, maintains competitive predictive performance (MAE) relative to dense baselines, and achieves substantially higher participation fairness than the energy-only policy. Ablation studies confirm the importance of fairness: removing the participation penalty reduces fairness to the level of energy-only scheduling. These findings support the central claim that urban FL should balance communication and energy efficiency without systematically excluding spatially distributed clients.

Instituciones: Universidad Nacional de San Agustín de Arequipa