A Federated Learning Framework for Inter-cell Handover Optimization in O-RAN

Efficient handover (HO) management remains a key challenge in 5G open radio access network (O-RAN) systems, particularly under dynamic traffic and mobility. Traditional A2/A3-based methods rely on static thresholds, causing handover failures, ping-pong effects, and load imbalance. While centralized machine learning (ML) improves decision-making, it introduces latency and signaling overhead. This study presents FLeXHO, a federated learning (FL)-based framework for mobility-aware inter-cell HO optimization within a near-real-time O-RAN architecture. The framework integrates three xApps - Metrics Collector, FL Predictor, and HO Handler - on a Flex RAN intelligent controller (RIC). Evaluation is conducted on a testbed with OpenAirInterface (OAI) and Open5GS, emulating 20 mobile users under a Gauss-Markov mobility model with variable traffic through iPerf. Results show over 94% HO success, significant load balancing gains, and reduced latency and packet loss, confirming the framework's effectiveness for real-time, distributed HO control in O-RAN deployments.