Treatment of Class Imbalance Problem in Noisy Dataset using SMART-SMOTE

Class imbalance poses a significant challenge in machine learning, often leading to biased models. Traditional oversampling techniques like SMOTE, ADASYN, and Borderline-SMOTE address this issue but face limitations such as presence of noisy points, poor neighbor selection, and irrelevant feature inclusion. This paper introduces SMART-SMOTE, a novel approach that mitigates these issues by removing noisy points, selecting relevant features, and using a dissimilarity-based neighbour selection mechanism. Gaussian noise is added to enhance robustness and generalization. Experiments on the Statlog German Credit and Online Shoppers’ Purchasing Intention datasets show that SMART-SMOTE outperforms state-of-the-art methods in accuracy and F1 score, effectively addressing class imbalance across diverse datasets and real-world applications like fraud detection and medical diagnosis.