ADAPTIVE FEATURE EXTRACTION AND COST-SENSITIVE NEURAL LEARNING FOR ROBUST CLASSIFICATION IN HIGH-IMBALANCE DOMAINS: BRIDGING BLIND SOURCE SEPARATION AND FRAUD DETECTION
Dr. Elena R. Vostik · Ethiopian International Journal of Multidisciplinary Research · 2025
The proliferation of high-dimensional datasets has exacerbated the challenge of detecting rare but critical events, such as financial fraud or distinct medical pathologies. Traditional machine learning algorithms, optimized for overall accuracy, frequently exhibit a bias toward the majority class, rendering them ineffective in high-imbalance scenarios. This article proposes and evaluates a novel hybrid architecture: the Independent Component Analysis–Cost-Sensitive Neural Network (ICA-CSNN). By integrating the blind source separation capabilities of ICA with the decision-boundary flexibility of neural networks, we aim to isolate latent "fraud" signals from background noise before classification. We first employ Information-Maximization (Infomax) ICA to transform correlated input features into statistically independent components, effectively disentangling the underlying data structure. Subsequently, these components serve as inputs to a Multi-Layer Perceptron (MLP) trained using a cost-sensitive backpropagation algorithm that disproportionately penalizes minority-class errors. Drawing upon seminal work in signal processing and recent advances in imbalance learning, we demonstrate that this dual-stage approach significantly improves sensitivity and F-measure scores compared to traditional methods. Our analysis suggests that enforcing statistical independence in the feature space aids the neural network in converging upon optimal decision boundaries, even when the minority class represents less than 1% of the data. The findings have direct implications for the design of robust automated detection systems in cybersecurity and healthcare.