Problem

Managed Wi-Fi networks increasingly utilize Distributed AI to predict Access Point (AP) operational statistics and optimize network performance. Federated Learning (FL) is well-suited as it avoids centralizing raw data. However, a single global FL model often performs suboptimally in managed Wi-Fi due to heterogeneous data distributions across APs. Clustered Federated Learning (CFL) addresses this by grouping similar APs and learning per-cluster models, but identifying informative and effective AP clusters from federated updates remains a significant challenge. Existing CFL efforts in Wi-Fi have not specifically targeted traffic prediction, and prior clustering in managed Wi-Fi typically reintroduces data centralization.

Method

The proposed Clustered Federated Learning with 2 Stage clustering (CFL-2S) method aims to generate more informative clusters for traffic prediction. In Stage 1, it first reduces the dimensionality of aggregated AP gradient features using Singular Value Decomposition (SVD) and then generates a pool of candidate partitions using k-means, varying the number of clusters. These candidates are filtered based on quality criteria, such as a minimum silhouette score, to ensure well-separated and cohesive clusters. If no satisfactory partitions are found, the system reverts to a single global FL model. In Stage 2, if candidate partitions exist, the method selects the final partition by maximizing the cumulative Gaussian differential entropy of its smallest cluster. This approach prioritizes partitions where the smallest, and thus most data-scarce, cluster contains highly informative gradient profiles, thereby reducing the risk of under-trained models for these groups. The method operates solely on gradient updates, making it agnostic to the learning task and data modality.

Results

CFL-2S was validated on a real-world campus Wi-Fi dataset for traffic prediction. It consistently achieved the best Mean Absolute Error (MAE) among all evaluated distributed strategies, including vanilla FL, IFCA, and CFL-GP. Specifically, CFL-2S reduced MAE by up to 51% over vanilla FL and 21% over CFL-GP, with the most significant improvements observed for longer prediction horizons (e.g., 1-hour ahead) and larger networks. Among clustered methods, CFL-2S also demonstrated the lowest communication overhead, exchanging approximately 73% less data than IFCA and 20% less than CFL-GP. Similarly, it consumed the least computational energy, averaging around 49% less than IFCA and 37% less than CFL-GP. While vanilla FL sometimes converged faster and had lower overall communication/energy costs for large networks at longer horizons (due to early stopping), CFL-2S maintained its lead in predictive accuracy.

Implications

The CFL-2S methodology provides a robust solution for traffic load prediction in centrally managed Wi-Fi networks, effectively addressing the challenges posed by heterogeneous AP data distributions. Its ability to generate informative clusters leads to significantly improved predictive performance compared to existing federated learning approaches, while maintaining efficient communication and energy consumption among clustered methods. The results suggest that clustered learning is most beneficial when AP data heterogeneity is high, justifying the additional overhead for substantial accuracy gains. Future work includes studying hyperparameter sensitivity, exploring alternative clustering quality criteria across diverse tasks, and developing cost-aware selective clustering to activate per-cluster models only when accuracy gains warrant the increased resource expenditure.