If Not Here, There. Explaining Machine Learning Models for Fault Localization in Optical Networks

Machine Learning (ML) is being widely investigated to automate safety-critical tasks in optical-network management. However, in some cases, decisions taken by ML models are hard to interpret, motivate and trust, and this lack of explainability complicates ML adoption in network management. The rising field of Explainable Artificial Intelligence (XAI) tries to uncover the reasoning behind the decision-making of complex ML models, offering end-users a stronger sense of trust towards ML-Automated decisions. In this paper we showcase an application of XAI, focusing on fault localization, and analyze the reasoning of the ML model, trained on real Optical Signal-To-Noise Ratio measurements, in two scenarios. In the first scenario we use measurements from a single monitor at the receiver, while in the second we also use measurements from multiple monitors along the path. With XAI, we show that additional monitors allow network operators to better understand model's behavior, making ML model more trustable and, hence, more practically adoptable.