Classifier-Agnostic Ambiguity Detection Framework for Targeted Oversampling in Imbalanced Learning

Abstract

This paper introduces a classifier-agnostic framework for ambiguity detection in imbalanced datasets, designed to improve classification performance by isolating regions of high local uncertainty. Unlike traditional resampling methods that treat the entire dataset uniformly, our framework identifies ambiguous regions based on misclassified minority instances and quantifies local uncertainty using entropy estimated over k-nearest neighborhoods. A distribution-driven threshold selection mechanism dynamically selects high-entropy zones that indicate structurally complex and failure-prone areas in the feature space. These regions are further aggregated using clustering techniques to ensure a coherent spatial representation. The resulting ambiguous zones can guide targeted oversampling strategies that enhance model performance while avoiding overfitting in well-separated regions. The framework is modular and classifier-independent, enabling systematic comparison across different models. Extensive experiments on 34 benchmark datasets demonstrate that ambiguity-guided interventions significantly improve recall, F1-score, and G-mean compared to global SMOTE approaches. The proposed method offers a statistically grounded, scalable, and interpretable approach to localized resampling, making it suitable for deployment in high-stakes domains such as medical diagnosis and anomaly detection.