Federated Aggregation Strategies Handover Prediction under Non-IID data conditions

Abstract

Decentralized Machine Learning (ML) through Federated Learning (FL) offers an alternative to centralized machine learning while preserving privacy for handover prediction, but client data distributions in networks such as fifth-generation (5G) are non-Independent and Identically Distributed (non-IID), which influences the choice of aggregation method needed for efficient results. This paper presents a comparative analysis of four FL aggregation strategies: FedAvg (Federated Averaging), FedProx (Federated Proximal), FedNova (Federated Normalized Averaging) and FedAdam (Federated Adaptive Adam Optimizer) that are integrated with an ExtraTrees ensemble classifier for proactive handover prediction. The evaluation was conducted under non-IID heterogeneity at various federation scales. In the reported partition, FedAdam achieves the highest accuracy of 95.3% at 5 clients but underperforms from 8 to 12 clients, where sample-weighted strategies lead; FedNova obtains the highest accuracy at 10 clients (94.5%) but degrades at 15 clients (91.7%, MCC = 0.894). A robustness check across three random seeds confirms the 15-client FedNova degradation as fully reproducible, while the specific strategy orderings at 5–12 clients fall within one standard deviation of run-to-run variability. Privacy costs increase from 2.70% to 5.30% with the number of clients and remain within acceptable bounds.