FedCORA: An interpretable context-aware aggregation framework for federated learning under real-world client heterogeneity

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Federated learning trains a shared model across clients that never pool their data, but accuracy degrades when clients are non-IID, and under the configurations evaluated here the standard correction (FedProx) did not improve on plain averaging. We propose FedCORA, an aggregation rule that reads each client’s operating context through a small interpretable fuzzy rule base. Four indicators computed from a client’s own record—a load indicator, an exceedance indicator, a signal-variability index and its recent validation error—set its aggregation weight, and the proximal pull is applied only to clients a stability gate flags unstable. We evaluate on a real heterogeneous problem: 60-minute-ahead indoor CO 2 forecasting across three operational schools whose distributions differ sharply (Jensen–Shannon divergence up to 0.24), using a low-cost sensor node we designed and deployed for two to three months per school; the federated protocol is emulated on the centrally collected datasets rather than run on site. Over ten independent runs, FedCORA reaches 178.6 ± 2.0 ppm RMSE against 181.2 ± 7.6 ppm for centralized training, a difference that is not statistically significant (p = 0.31) and does not establish equivalence, and it significantly outperforms FedAvg, equal-weight averaging, FedProx, SCAFFOLD, and FedNova (all p ≤ 0.012). By tying aggregation to context indicators rather than data volume, FedCORA gives heterogeneous sensor networks an aggregation rule that is stated explicitly rather than learned; a synthetic federation of up to 40 clients shows the same hand-set weighting continues to run without destabilizing.

• FedLucra: Federated Learning With Loss‐Utility Controlled Robust Aggregation

  Expert Systems 2026 · 0 citations