Enhancing Reliability of Lightpath QoT Estimation Models using Bias Mitigation Techniques

Quality of Transmission (QoT) classification is a critical task in optical networks, often managed via Machine Learning (ML) models. Traditional ML models may exhibit unintended bias toward certain system parameters. In this paper, we explore multiple strategies to mitigate bias toward modulation formats in a large-scale QoT dataset containing more than 1.3 million samples and 32 features. We systematically proposed a hybrid approach that combines three techniques: (i) preprocessing, (ii) in-processing, and (iii) post-processing.We employ XGBoost as our baseline classifier and measure performance using accuracy, ROC-AUC, precision, and recall, along with fairness metrics such as Demographic Parity Difference, Equalized Odds Difference, and Predictive Value Parity. Our results demonstrate that hybrid solutions substantially reduce bias while preserving competitive model performance. These findings highlight the feasibility of combining complementary fairness approaches to achieve fair QoT predictions without a significant loss in classification accuracy. We discuss practical considerations for deploying such techniques in real-world systems and outline directions for future work, including extended subgroup analysis and alternative pre-processing methods.