Reducing Complexity and Enhancing Predictive Power of ML-Based Lightpath QoT Estimation via SHAP-Assisted Feature Selection

In dynamic optical network planning, accurate estimation of the Quality-of-Transmission (QoT) using Machine Learning (ML) for online lightpaths (LPs) is crucial. While existing research has focused extensively on improving the precision of ML-based QoT estimation models, there has been limited attention given to factors such as model complexity, inference speed and the telemetry load associated with the ML pipeline. These factors are pivotal for real-time deployment, as complex models can strain computational resources and slow down decision-making during LP provisioning. Hence, a balance between predictive accuracy, computational efficiency, telemetry load, and inference speed is essential for the development of sustainable and lightweight QoT estimation solutions. In this work, we first compare how different ML methods perform in LP QoT estimation in terms of predictive power and computational complexity. We show that Deep Neural Network (DNN)-based models, which are widely applied in literature, are costly in terms of computational requirements and have an inference speed that may restrain practical deployments, even when applying knowledge distillation. We then exploit Shapley Additive Explanations with the aim of determining and selecting the feature set that most impacts the model's outcome. Numerical results quantify the trade-off between predictive power and computational capacity and show that the feature set can be reduced, reducing telemetry load and improving inference speed, while improving the model's predictive accuracy. Quantifying the trade-off offers guidance to operators who wish to utilize lightweight models for estimating LP QoT without compromising on performance.