Aladin, Sandra, Wosinska, Lena et Tremblay, Christine.
2025.
« Automated, interpretable and efficient ML models for real-world lightpaths’ quality of transmission estimation ».
IEEE Open Journal of the Communications Society.
(Sous presse)
Prévisualisation |
PDF
Tremblay-C-2025-33029.pdf - Version publiée Licence d'utilisation : Creative Commons CC BY. Télécharger (3MB) | Prévisualisation |
Résumé
Fast and accurate estimation of lightpaths’ quality of transmission (QoT) is crucial for ensuring quality of service (QoS) and seamless operation in real-world optical networks. Machine learning (ML) algorithms are promising tools for QoT estimation of lightpaths before their establishment. In multidomain optical networks, where learned QoT estimation models must be transferred between heterogeneous environments with limited target data, deep neural networks (DNNs) have shown promising results. However, DNN-based transfer learning (TL) approaches using fine-tuned artificial neural networks (ANNs) and convolutional neural networks (CNNs), offer limited interpretability. Consequently, little insight into the decision-making process is provided, and large labeled datasets as well as high computational resources are required, limiting their suitability for real-time, large-scale deployment in production networks. To address these challenges, we propose a novel lightweight and interpretable TL framework that integrates the Boruta-SHAP algorithm for automated feature selection (FS) together with two domain adaptation (DA) techniques: Feature Augmentation and Correlation Alignment. In contrast to the existing approaches based on DNN, our strategy leverages interpretable and efficient ML models to enhance the adaptability across diverse datasets in real-world network environments. We show that our random forest (RF)-based models achieve better performance than the ANN-based models, without sacrificing the classification accuracy. The FS via Boruta-SHAP allows for reducing dimensionality as well as training and inference times up to 70.68%, and 41.64%, respectively. Our proposed framework outperforms DA baseline models achieving 99.35% accuracy improvement in domain shift. Moreover, it offers 86% accuracy with a 99.83% reduction in the size of the target domain.
| Type de document: | Article publié dans une revue, révisé par les pairs |
|---|---|
| Professeur: | Professeur Tremblay, Christine |
| Affiliation: | Génie électrique |
| Date de dépôt: | 21 nov. 2025 20:56 |
| Dernière modification: | 28 nov. 2025 17:55 |
| URI: | https://espace2.etsmtl.ca/id/eprint/33029 |
Actions (Authentification requise)
 |
Dernière vérification avant le dépôt |