Hussein Fawaz, Farhad Arpanaei, Davide Andreoletti, Ihab Sbeity, José A. Hernández, David Larrabeiti, Omran Ayoub
Lebanese University, SUPSI
IEEE ONDM 2024
Other version: [PDF] [[Code – forthcoming]]
TL;DR
We use SHAP-based feature selection to build simpler and more accurate ML models for lightpath quality-of-transmission (QoT) estimation in optical networks.
The complexity problem in QoT estimation
QoT estimation models often rely on large feature sets, increasing computational cost and limiting interpretability—both problematic in operational optical networks.
SHAP-assisted feature selection
SHAP values are used to quantify feature importance and systematically remove redundant inputs. This leads to:
- Reduced model complexity
- Faster inference
- Improved interpretability
Results
Our method achieves:
- Higher predictive accuracy
- Smaller feature sets
- Better generalization
Broader impact
This work demonstrates how explainable AI can be leveraged not only for interpretation, but also for practical system optimization.
Future work
We aim to apply this framework to other networking problems and integrate reliability-aware feature selection.