Risk-Aware Machine Learning-Based Approach for Lightpath QoT Estimation in Optical Networks

Machine Learning (ML)-based models for Quality of Transmission (QoT) estimation in optical networks are commonly trained and evaluated using aggregate accuracy metrics such as Mean Squared Error (MSE), which fail to reflect the operational impact of prediction errors. In practice, QoT predictions are compared against feasibility thresholds to make accept/reject de cisions, causing errors near such thresholds to entail significantly higher operational risk than equally sized errors far from them. To address this asymmetry in terms of operational risk, we introduce a risk-aware ML-based QoT estimation framework based on a Hybrid Ensemble (HE) architecture, which partitions lightpaths by ranges of Signal to Noise Ratio (SNR) margins and allocates higher modeling capacity to operationally critical regions of the feature space. We also propose metrics to quantify expected decision risk and risk disparity across lightpath groups. Experiments on two Space Division Multiplexed (SDM) network datasets show that the proposed approach consistently reduces both aggregate and group-level decision risk. Compared to the best-performing baseline, it achieves up to 7% reduction in prediction error, 8% reduction in expected decision risk, and 7% reduction in risk disparity, while maintaining low system complexity.