Preprint: Measuring Concept Completeness for Explainability in AI-Driven Network Functions

Abstract

Concept-based explanations ground ML decisions in domain concepts, assuming a concept set that captures the model’s reasoning without verification. We introduce a representation-level framework for quantifying concept completeness, measuring the variance in internal representations captured by the concept subspace, yielding a global metric κ∗ ∈ [0, 1]. For high-dimensional, noisy spaces, we add a significant-component measure τ∗ via singular value decomposition and a Shapley-based importance metric. On three AI networking scenarios, we find congestion control substantially complete and the other two only moderately complete. We presented this work at the IFIP Networking 2026 conference.