Natural Language Interpretability for ML-Based QoT Estimation via Large Language Models

As Machine Learning (ML) systems become integral to network management, the need for transparent decision-making grows. While post-hoc explainability methods provide insights into model behavior, their technical nature often limits accessibility. We explore Large Language Models (LLMs) for translating complex ML model explanations, extracted using explainable artificial intelligence frameworks, into natural language to simplify user understanding and interpretability. Using direct prompting and self-reflection-based prompting, we generate explanations for a lightpath Quality of Transmission (QoT) estimation model. Empirical evaluations confirm the correctness and usefulness of LLM-generated interpretations in about 65% of the cases, highlighting the benefits of self-reflection in enhancing explanation quality. The study also remarks on the necessity of devising enhancements to improve the results achieved so far.