As fifth-generation (5G) networks grow in complexity and increasingly incorporate AI-driven autonomous management, the ability to explain automated decisions has become critical for operator trust, regulatory compliance, and network governance. Explainable Artificial Intelligence (XAI) methods, and in particular SHAP (SHapley Additive exPlanations), provide principled feature attribution scores for individual predictive models. However, in multi-domain 5G architectures where distinct machine-learning models predict Quality of Service (QoS) KPIs across the Radio Access Network (RAN), Core, and Transport domains, explanations are inherently model-specific, operate over partially heterogeneous feature subsets, and are of unequal reliability. While prior work has proposed aggregation methods for feature-based explanations, no existing framework simultaneously addresses quality-informed weighting, formal aggregation across heterogeneous feature spaces, and alignment with operator-facing intents in a multi-domain 5G setting.
This thesis proposes a reliability-aware framework for aggregating SHAP feature attributions across heterogeneous 5G domain models. Central to the framework is the Explanation Aggregation Metric (EAM), which scores each model’s explanation quality along three dimensions: Fidelity, Stability, and Consistency. These scores drive an EAM-weighted convex aggregation along the cross-KPI axis, which combines the different-KPI models of a single domain into a domain-level attribution profile. Along the complementary cross-domain axis, the per-domain models of a single KPI are instead composed through that KPI’s physical QoS operator (additive for latency, minimum for throughput, and product for availability) rather than by reliability weighting, since the domains are related by a known composition law; this composition layer produces intent-aligned end-to-end explanations.
Evaluated on gradient-boosted tree models trained across three network domains and three KPIs, EAM weighting attenuates contributions from less reliable models in proportion to their measured explanation quality, without discarding them entirely, and top-ranked features remain consistent across both weighting schemes, confirming that dominant signals are robust to the choice of aggregation. Cross-model-family agreement on feature rankings within each domain, together with cross-domain agreement on the physical relevance of load-related features, provides empirical evidence that the framework recovers genuine network structure. The thesis identifies limitations of the current EAM formulation, most notably the absence of range-normalisation in the Stability score and the sampling-based nature of KernelSHAP attributions for composite operators, and outlines probable future works.