Monotone FedNAM enforces physics in O-RAN SLA risk predictions, cuts traffic 65%

Proactive service assurance in O-RAN (Open Radio Access Network) requires predicting per-slice SLA violations before they occur, but models must be auditable and train across base stations without pooling commercially sensitive per-slice KPIs. Neural additive models (NAMs) offer auditability through visible shape functions, but unconstrained NAMs learn effects that contradict wireless physics. On the ColO-RAN testbed, these models predicted higher risk when channel quality improved — a clear physical impossibility. This failure occurred under both local and centralized training and worsened with non-IID federated averaging.

To solve this, the authors introduce Monotone FedNAM, a federated additive model where KPIs with unambiguous physical direction are represented as monotone splines whose constraints survive FedAvg aggregation by construction, while contestable KPIs remain unconstrained. The model operates as both a Non-RT RIC rApp and a compact Near-RT RIC xApp. Results show it eliminates all monotonicity violations, raises constrained shape consistency from 0.71 to 1.00, reduces uplink traffic by 65%, and generalizes to an unseen scheduling policy — at a cost of only 0.04 to 0.07 AUC. This demonstrates that physically constrained federated additive models can support auditable, physically valid SLA risk inference for multi-tenant O-RAN service assurance.