Causal Explanations for Performance in Radio Networks

Domokos M. Kelen, Péter Kersch, András A. BenczúrAI6G@WCCI 2022 · OpenReview

Introduction

As mobile networks transition toward 5G and 6G, the complexity of managing radio access networks (RAN) has increased exponentially. While Artificial Intelligence (AI) is often proposed as the solution for automated network management, many existing models operate as "black boxes." When performance metrics like throughput degrade, standard AI models can identify correlations—such as high interference—but often fail to pinpoint the actionable root cause, such as an incorrect antenna tilt or cell configuration. This paper introduces a novel approach to Explainable AI (XAI) that incorporates causal domain knowledge to provide engineers with actionable insights.

The Problem: The "Mediator Trap"

Standard XAI methods, most notably SHAP (Shapley Additive Explanations), are designed to attribute importance based on statistical correlation. In a radio network, features exist in a hierarchy: configuration settings affect intermediate metrics (mediators) like coverage or interference, which in turn determine final performance (throughput).

Because intermediate metrics are mathematically closer to the target result, standard models are "lazy" and attribute the majority of the importance to these mediators. For a network engineer, being told that "low throughput is caused by low channel quality" is a tautology; they need to know which configuration setting caused the quality to drop in the first place. This gap between correlation and causation is the primary hurdle for effective network automation.

Methodology: Respecting Causal Order

To address this, the researchers propose forcing XAI to respect a pre-defined "causal order" of features. This requires human domain expertise to construct a causal graph before analysis begins.

Figure 1: A causal graph illustrating the relationships between cell configuration, intermediate metrics like interference and cell load, and the final throughput target.

The authors developed two primary techniques to implement this causal constraint:

  1. Causal-Order Training: A Gradient Boosted Decision Tree (GBDT) is trained in stages. The model is first exposed only to "root cause" features (e.g., configuration). Only after learning these relationships is it allowed to see intermediate features to refine its predictions. This forces the model to prioritize causal ancestors.
  2. Conditional Asymmetric SHAP (Separate Models): This more robust approach involves training a separate model for every valid combination (coalition) of feature groups allowed by the causal graph. By comparing models with and without specific features, the system can precisely isolate the unique causal contribution of a feature, effectively bypassing the statistical noise of mediators.

Experimental Results

Synthetic Validation

The methods were first tested on a synthetic dataset where the ground truth was known: variables A, B, and C caused D, which then determined the target Y.

Figure 2: Comparison of feature importance (Shapley values) across different models. Note how the 'Separate models' approach correctly attributes zero importance to the mediator (D).

As shown in the results, a regular model attributed nearly 80% of the importance (0.7928) to the mediator (D). In contrast, the "Separate Models" approach successfully shifted this importance back to the actual causes (A, B, and C), assigning the mediator a value of 0.0000.

Real-World Telecom Application

When applied to proprietary Performance Management (PM) data from live radio cells, the "Separate Models" approach significantly outperformed standard XAI. It highlighted the impact of "TA Distribution" (representing user distance and location) as a primary driver of performance. Standard models, conversely, over-emphasized "Cell Load," which is often a symptom rather than a root cause of performance shifts.

Conclusion and Future Work

The study demonstrates that domain knowledge is mandatory for meaningful XAI in complex systems. You cannot infer causality from raw data alone; the model must be guided by a human-defined causal structure. While the "Separate Models" approach is computationally intensive, it provides the level of accuracy required for "closed-loop" automation, where AI can not only explain a problem but also execute the correct configuration change to fix it. Future research will focus on reducing the computational overhead of training multiple coalition models to make this approach scalable for networks with hundreds of variables.

Summary updated

BibTeX
@inproceedings{kelen2022causal,
  author = {Domokos M. Kelen and Péter Kersch and András A. Benczúr},
  title = {{Causal Explanations for Performance in Radio Networks}},
  booktitle = {AI6G@WCCI},
  year = {2022},
  url = {https://openreview.net/forum?id=RpnkAvtpSN},
}
Figure