Generative Explainability for Next-Generation Networks: LLM-Augmented XAI with Mutual Feature Interactions

Researchers present an LLM-augmented explainable AI framework that generates human-readable explanations for network operations by combining SHAP feature analysis with mutual feature interactions. The approach demonstrates 12.2% improvement in explanation usefulness over baseline methods while maintaining 97.5% correctness, addressing the critical gap between opaque AI/ML models and operator trust in network infrastructure.

The growing reliance on AI/ML models within critical network infrastructure creates a fundamental trust problem: operators cannot confidently deploy systems they cannot understand. This research tackles a genuine pain point in enterprise AI adoption by moving beyond purely technical explanations toward human-comprehensible narratives. The framework bridges machine learning interpretability and natural language generation, two increasingly mature domains, to produce actionable insights for non-specialist stakeholders.

The approach builds on established XAI methodology—specifically SHAP values, which quantify feature importance—but enriches it with mutual feature interaction data and LLM-based natural language translation. This represents an incremental but meaningful advancement in making AI systems more accessible to practitioners who lack deep technical backgrounds. The empirical validation through human evaluators and measurement of inter-evaluator agreement strengthens the credibility of the results.

For enterprise network operators and AI system implementers, this framework has practical implications. Better explanations reduce deployment friction, shorten evaluation cycles, and enable faster identification of model failures or drift. The 12.2% improvement in usefulness suggests tangible value in production environments. The optical QoT estimation use case demonstrates applicability to real infrastructure challenges, particularly relevant for telecommunications and internet backbone operators managing increasingly complex automated systems.

Looking forward, the critical question is scalability: whether moderately-sized LLMs can maintain explanation quality across diverse network scenarios and use cases. Further research on computational overhead, latency impacts on real-time network operations, and applicability beyond optical networks will determine whether this approach becomes standard practice in network AI deployment.

• →LLM-augmented XAI framework improves explanation usefulness by 12.2% compared to SHAP-only baselines while maintaining 97.5% correctness.
• →Integration of mutual feature interactions with natural language generation bridges the gap between technical AI outputs and operator-understandable insights.
• →Human evaluator testing with high inter-rater agreement validates the practical utility of generated explanations for non-specialists.
• →Framework demonstrates applicability to critical infrastructure (optical QoT estimation) where transparency directly impacts deployment confidence.
• →Research addresses enterprise AI adoption barrier by making opaque model decisions accessible without requiring specialist interpretation skills.