⚡ electrical engineering

Physically Constrained Federated Additive Models for O-RAN SLA-Risk Prediction

This paper introduces Monotone FedNAM, a physically constrained federated additive model that ensures auditable and physically valid SLA-risk predictions for multi-tenant O-RAN by enforcing monotonicity on key performance indicators during federated training, thereby eliminating contradictory predictions while significantly reducing uplink traffic with minimal accuracy loss.

Original authors: Aubida A. Al-Hameed, Mohammed M. H. Qazzaz, Maryam Hafeez, Syed A. Zaidi

Published 2026-07-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aubida A. Al-Hameed, Mohammed M. H. Qazzaz, Maryam Hafeez, Syed A. Zaidi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the conductor of a massive, high-speed orchestra where every musician is playing a different instrument, but they are all in separate soundproof rooms. You need to know if the music is about to go off-key before it happens, so you can fix it instantly. This is the challenge of modern mobile networks, specifically a new, flexible system called O-RAN. In this world, the network is split into "slices," like separate lanes on a highway, each rented out to different companies (tenants) who have their own strict rules about speed and reliability. The problem is that these companies don't want to share their private traffic data with each other or the network operator because it's a trade secret.

To solve this, engineers use a technique called "Federated Learning." Think of it like a group of students working on a math problem. Instead of sending their private notebooks to a central teacher, they each solve their part of the problem in their own room and only send the final answer (the "model update") to the teacher. The teacher combines these answers to create a master guide. However, there's a catch: sometimes the students' answers are so weird or contradictory that the master guide makes no sense. For example, the guide might suggest that "better weather" causes "more traffic jams," which is physically impossible. This paper tackles the problem of making sure these AI guides not only keep secrets but also follow the basic laws of physics and logic, ensuring the network doesn't give dangerous advice.

The Paper's Story: Teaching AI to Respect Physics

The researchers behind this paper, working with data from a real-world testbed in Rome, discovered a surprising flaw in how AI predicts network failures. They were using a type of smart model called a "Neural Additive Model" (NAM). You can think of a NAM as a team of specialists, where each specialist is responsible for just one factor, like "how many people are online" or "how strong the signal is." Because each specialist works alone, you can look at their work and see exactly how they think. This is great for "auditability"—it means a human can check the math.

But here's the twist: just because you can see the math doesn't mean it's right. The team found that when these models learned from data, they often came up with conclusions that defied common sense. In one case, the model learned that when the radio signal got better, the risk of the network failing actually went up. It's like a weather app telling you that a sunny day is a sign to grab an umbrella because a storm is coming. This happened even when the models were trained in the center, and it got even worse when the models tried to learn together from different locations (federated learning). The different "voices" of the base stations were mixing together to create a distorted, physically impossible global rule.

The Solution: The "Monotone" Guardrails

To fix this, the authors created a new system called Monotone FedNAM. Imagine you are training a robot to drive a car. You don't just let it learn by watching; you install hard-coded guardrails. If the robot sees a red light, the guardrail forces it to stop. If it sees a green light, the guardrail forces it to go. You can't tell the robot to stop at green or go at red, no matter how much data it sees.

In this paper, the "guardrails" are mathematical constraints called monotonicity. The researchers identified which network factors have a clear, unambiguous direction. For example, if "channel quality" (signal strength) goes up, the risk of failure must go down. If "backlog" (a pile-up of data) goes up, the risk must go up. They built the model so that for these specific factors, the AI is physically forced to follow these rules. It's like telling the AI, "You can learn anything you want about the traffic, but you are strictly forbidden from thinking that a stronger signal causes more crashes."

What They Found

When they tested this new system on the ColO-RAN testbed (which simulated seven different base stations), the results were striking:

Zero Physics Violations: The old, unconstrained models broke the laws of physics constantly. The new Monotone FedNAM broke them zero times. It completely eliminated the weird, inverted predictions.
The Cost of Safety: There was a tiny price to pay for this safety. The model's overall prediction accuracy (measured by a score called AUC) dropped slightly, from about 0.943 (for the unconstrained version) to 0.901 (for the constrained version). The authors note this is a small trade-off: you lose a tiny bit of raw guessing power to gain total trust in the logic.
Better Communication: Because the model had to follow strict rules, it didn't need to send as much data back and forth between the base stations and the central server. The amount of data sent over the network dropped by 65%, making the system much faster and cheaper to run.
Generalization: The model was able to take what it learned and apply it to a new, unseen scheduling policy (a new way of managing traffic) without breaking down, proving it learned the rules of the road, not just memorized the traffic patterns.

The Big Picture

The paper concludes that while "black box" AI models might be slightly more accurate at guessing numbers, they are dangerous for critical infrastructure because they can't explain why they made a mistake, or worse, they might make a mistake that looks logical but is physically impossible. By forcing the AI to respect the known laws of physics (like "better signal = lower risk"), the researchers created a system that is not only accurate enough to be useful but also safe enough to be trusted by network operators. They showed that you can have an AI that is both a smart predictor and a logical thinker, provided you give it the right guardrails to work within.

Technical Summary: Physically Constrained Federated Additive Models for O-RAN SLA-Risk Prediction

Problem Statement

In Open Radio Access Networks (O-RAN), proactive service assurance requires predicting Service Level Agreement (SLA) violations for individual network slices before they occur. This task faces three critical challenges:

Data Privacy and Sensitivity: Per-slice Key Performance Indicators (KPIs) reveal tenant traffic patterns and load, making them commercially sensitive. Pooling raw data across base stations is often contractually or legally prohibited, necessitating a federated learning (FL) approach where models train locally and only updates are shared.
Auditability and Trust: Operators must be able to audit ML models before they drive control loops. Standard deep learning models (e.g., deep reinforcement learning) are opaque "black boxes," and post-hoc explainability methods are often unfaithful or computationally prohibitive for Near-Real-Time (Near-RT) RIC inference.
Physical Validity: While Neural Additive Models (NAMs) offer intrinsic interpretability by learning a visible shape function for each feature, the authors demonstrate that interpretability does not guarantee physical validity. On measured O-RAN traces, unconstrained NAMs frequently learn effects that contradict wireless physics (e.g., predicting higher SLA risk when channel quality improves). This issue is exacerbated in federated settings where non-IID (non-independent and identically distributed) data across clients leads to distorted global effects upon aggregation.

Methodology: Monotone FedNAM

The authors propose Monotone FedNAM, a federated neural additive model designed to enforce physical constraints while maintaining interpretability and privacy.

Model Architecture: The model decomposes the prediction into a sum of feature-specific functions ( $f(x) = \beta_0 + \sum g_j(x_j)$ $f (x) = β_{0} + \sum g_{j} (x_{j})$ ).
- Constrained Features: KPIs with unambiguous physical directions (e.g., higher backlog increases risk; better channel quality decreases risk) are represented as monotone linear splines. These splines use a shared grid of 16 knots placed at marginal feature quantiles. The shape is enforced by parameterizing the spline increments using a softplus function, ensuring all increments are strictly positive (or negative), thereby guaranteeing monotonicity by construction.
- Unconstrained Features: KPIs with contestable or unclear physical directions (e.g., specific allocation metrics) are modeled using unconstrained small Multi-Layer Perceptron (MLP) subnetworks or linear terms.
Federated Aggregation with Constraints: The model employs a modified FedAvg algorithm. Crucially, the monotonicity constraints are preserved during aggregation. Because the constraints are encoded in the signed increments of the spline parameters, a weighted average of non-negative (or non-positive) increments from local clients remains non-negative (or non-positive) at the server.
Anchoring for Identifiability: To prevent arbitrary vertical shifts in local shape functions from distorting the global model during aggregation, each feature function is anchored at a fixed public reference point (e.g., $x=0$ for standardized features). The offset is removed locally and folded into the global intercept.
Deployment Path: The model is trained and governed as a Non-RT RIC rApp (responsible for federated training and model governance). The resulting global model, containing parameters and a constraint manifest, is deployed as a Near-RT RIC xApp for online, per-slice SLA-risk inference with additive explanations.

Key Contributions

Demonstration of Physical Invalidity in Unconstrained NAMs: The authors show that on measured O-RAN traces, unconstrained additive models learn physically incorrect shape functions (e.g., inverting the relationship between channel quality and risk). This failure occurs in both centralized and federated settings, with non-IID federated averaging worsening the distortion.
Monotone FedNAM: A novel federated architecture that preserves monotonicity under aggregation by construction. It selectively constrains only KPIs with unambiguous physical directions while leaving contestable KPIs unconstrained, balancing physical validity with model flexibility.
rApp-to-xApp Deployment Framework: The paper defines a practical deployment path where the Non-RT RIC governs the training of a physically consistent model, which is then deployed to the Near-RT RIC for auditable, real-time inference.

Experimental Results

The model was evaluated on the ColO-RAN testbed dataset, featuring seven base stations, three slice types (eMBB, MTC, URLLC), and three scheduling policies.

Physical Validity: Monotone FedNAM achieved a monotonicity violation rate of 0.000 on the constrained KPI set. In contrast, unconstrained local NAMs had a violation rate of 0.461, centralized NAMs 0.508, and unconstrained FedNAM 0.652. The constrained model raised the "constrained shape consistency" (agreement on the form of KPI effects across clients) from 0.71 to 1.00.
Predictive Performance: Enforcing physical constraints incurred a modest cost in accuracy. Monotone FedNAM achieved an AUC of 0.901, compared to 0.943 for unconstrained FedNAM and 0.973 for a black-box FedAvg-MLP. The cost of physical governance was approximately 0.04 AUC relative to the unconstrained federated model.
Efficiency and Generalization:
- Communication Overhead: By representing constrained features as 16-value increment vectors rather than hundreds of MLP weights, the model reduced the total uplink traffic by 65% (from ~10.75 MB to 2.15 MB over 30 rounds).
- Generalization: The model successfully generalized to an unseen scheduling policy (leave-one-scheduler-out test) with an AUC of 0.887, demonstrating that the governed KPI-risk relationships are robust.
- Leakage Robustness: Removing features prone to data leakage (e.g., current delivered rate) had only a minor impact on performance, confirming the model relies on genuine causal features.

Significance and Claims

The paper argues that interpretability alone is insufficient for O-RAN service assurance; models must also be physically valid to be trustworthy. The authors claim that:

Physical constraints must be explicit priors: Relying on data alone to learn correct wireless physics is insufficient, especially in federated settings with non-IID data.
Selective governance is optimal: Constraining only KPIs with unambiguous physical directions (the "conservative set") removes all violations and aligns client shapes without sacrificing the interpretability of contestable features. Over-constraining (the "aggressive set") reduces accuracy without adding interpretability value.
Feasibility for O-RAN: Physically constrained federated additive models can support auditable SLA-risk inference. The proposed Monotone FedNAM is compact enough for Near-RT RIC deployment and reduces communication overhead, making it a viable solution for multi-tenant O-RAN environments where data privacy and model trust are paramount.

The study acknowledges limitations, noting it was conducted on a single emulated testbed with static users and a proxy URLLC label, and future work is needed to validate these findings across dynamic mobility, larger federations, and multi-operator trust domains.