Skip to content
Open access

Imbalanced Learning for Multiclass Fault Prediction in Optical Networks Using Enhanced Hybrid Sampling

Oct 2026 · Emerging Science Journal · 0 citations · 41 references

Abstract

Fault prediction in optical networks faces the problem of class imbalance, where failure event data is much less than normal event data. Conventional classification models typically used in balanced or low-imbalance conditions tend to be biased toward the normal class. Sampling techniques such as oversampling tend to increase overlap. This causes high performance values in imbalanced classes not always to reflect the model's true ability to detect minority classes. This paper proposes a new framework called SMOTE-ENN-CatBoost-Optuna. This approach combines a hybrid sampling technique using synthetic minority oversampling (SMOTE) and edited nearest neighbors (ENN), controlled by Categorical Boosting (CatBoost), with hyperparameter tuning using Optuna. The proposed method dynamically adjusts the oversampling rate based on model error, followed by noise filtering to refine class boundaries. An iterative training strategy with Matthews correlation coefficient (MCC)-based convergence is further introduced to stabilize performance. To assess the proposed model's ability to adapt to varying levels of imbalance, imbalanced classes were grouped into three categories (medium, high, and extreme) based on the Imbalance Ratio (IR). Experimental results show that the proposed model consistently outperforms baseline methods. It achieves improvements of up to 0.9944 in MCC and 0.9951 in Macro F1-score. These results indicate the framework effectively enhances fault-detection reliability and maintains robust performance across most classes.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.