Federated Additive Models Predict O-RAN SLA Risk with Physics Constraints

Proactive management of service level agreements (SLAs) in Open Radio Access Networks (O-RAN) requires accurate prediction of potential violations. This is complicated by the need to train models across multiple base stations without centralizing commercially sensitive per-slice Key Performance Indicators (KPIs). Additionally, the prediction models must be auditable by operators, meaning their logic should be transparent. Neural Additive Models (NAMs) offer this transparency, but unconstrained NAMs can learn relationships that contradict fundamental wireless physics, such as predicting higher risk when channel quality improves. To address these issues, researchers developed Monotone FedNAM, a federated additive model. This model incorporates physical constraints by representing KPIs with clear physical directions (e.g., channel quality) as monotone splines. These constraints are designed to persist even after federated averaging, ensuring physical validity across the distributed network. Other, less clear-cut KPIs remain unconstrained. Evaluated on the ColO-RAN testbed, Monotone FedNAM successfully eliminated all monotonicity violations, achieving perfect constrained shape consistency. It also demonstrated improved generalization to unseen scheduling policies and reduced uplink traffic by 65%, with only a minor trade-off in AUC performance. This research shows that integrating physical constraints into federated additive models can create auditable and reliable SLA risk inference systems for multi-tenant O-RAN environments.