Explainable SHAP-based feature selection for accurate and computationally efficient network response time prediction in optical communication networks

Abstract This paper proposes an explainable SHAP-guided machine learning framework for optical network response-time prediction, in which SHapley Additive exPlanations (SHAP) are employed not only to interpret model behavior but also to identify and retain the most influential network parameters. A comprehensive set of conventional machine learning and ensemble regression models is evaluated under identical conditions before and after SHAP-based feature selection, including Random forest, gradient boosting, XGBoost, K-nearest neighbors, support vector regression, and an RF–GB–XGB stacking ensemble. The SHAP procedure reduces the transformed feature space while retaining the most informative network characteristics. Experimental results demonstrate that the effect of feature selection is model-dependent, with Random forest achieving the best overall post-SHAP predictive performance, obtaining an MAE of 405.2978, RMSE of 661.0859, and R 2 of 0.7196, corresponding to MAE and RMSE improvements of 3.90 % and 4.85 %, respectively. In addition to prediction accuracy, computational efficiency is assessed in terms of training time, RAM usage, inference latency, and model size. SHAP-based feature reduction generally decreases training cost, with Random forest reducing its training time by 12.79 %, while gradient boosting achieves a reduction of 51.44 %. Overall, the results demonstrate that SHAP can serve as an effective explainability-guided feature-selection mechanism, providing a practical balance among prediction accuracy, model interpretability, and computational efficiency for intelligent optical network performance management.