A comparative study on the noise sensitivity of binary classification based on robust deep neural networks

Khudhair, Mohammed Aad and Fazekas, Attila (2025) A comparative study on the noise sensitivity of binary classification based on robust deep neural networks. ANNALES MATHEMATICAE ET INFORMATICAE, 61. pp. 129-140. ISSN 1787-6117

Preview

Text 129_140_fazekas.pdf - Published Version Download (737kB) | Preview

Abstract

Tackling the persistent dual challenge of noise and class imbalance in binary classification, this study introduces a robust hybrid pipeline that improves resilience and accuracy in noisy, imbalanced data environments. Leveraging a multi-stage framework, we integrate a Gaussian Mixture Model Noise Filter (GMMNF) to preserve minority class integrity, a Noise- Aware Multi-Layer Perceptron (MLP) enhanced with dynamic regularization to adaptively mitigate noise, and a synergistic resampling strategy combining SMOTE-Tomek and Conditional GAN to optimize class distribution. Comprehensive evaluations across escalating noise levels (0–32%) reveal that our approach not only achieves a peak F1-score of 0.9255 at 4% noise but also maintains over 49% minority class representation even under severe noise stress. Five-fold cross-validation substantiates the pipeline’s robustness, consistently outperforming established state-of-the-art methods. These results underscore the significant advancement our framework offers for real-world applications where data imperfection and imbalance are the norm, in reliable binary classification.

Actions (login required)

Edit Item