A Hybrid Approach for Imbalanced Data Classification by Combining Clustering with Instance Selection

Class imbalance is characterized by an uneven distribution of classes, resulting in significantly more instances in one class (majority class or negative class) than in the others (minority class or positive class). Consequently, biased model performance may occur in favor of the majority class due to its higher prevalence, neglecting the minority class with fewer instances. This compromises the overall model performance, especially when accurately identifying the minority class is crucial. To solve this challenge, current methods balance the data distribution through oversampling, undersampling, or hybrid approaches. Our work presents a more efficient hybrid technique. We cluster minority instances utilizing k-means, obtain their densities inside each cluster, and produce new instances synthetically by choosing high-density minority instances. Then we cluster majority instances utilizing k-means, determine their densities inside each cluster, and discard less significant boundary-side majority instances that are close to the minority class. To obtain a balanced data distribution across classes, the high-density majority instances are chosen, and the remaining low-density instances are removed. We measure the performance of our proposed method using F1-Score and AUC and compare it with other established techniques on 10 imbalanced dataset demonstrating the superiority of our approach in both metrics.