Model and data-centric machine learning algorithms to address data scarcity for failure identification

Abstract

The uneven occurrence of certain types of failures in optical networks results in a scarcity of data for less frequent failures, leading to imbalanced datasets for training machine learning (ML) models. This poses a significant bottleneck in terms of reliability and practical implementation of ML for failure management. Existing research works often overlook this aspect while demonstrating high accuracies by utilizing sufficiently balanced training datasets collected in controlled laboratory setups and simulations. However, this approach does not reflect a realistic network scenario. To address this issue, different model-centric and data-centric approaches have been investigated in this work to determine their potential for improving the learning of ML models, specifically neural networks (NNs), on less frequent failures with such imbalanced training datasets. For failure identification, the obtained results suggest that data-centric approaches tend to perform better in terms of classification accuracy, with an improvement of up to 5.5% in F1-score observed on less frequent failures compared to a baseline NN (i.e., without any model-centric or data-centric treatment). However, some data-centric approaches may also have significant additional computational complexity associated with them, and, therefore, a suitable approach should be chosen based on the desired classification performance and available computational resources.