Borderline Active Learning: Transactional Records in Alert-Feedback System

Abstract

—Transactional records often exhibit highly imbalanced patterns, which can hinder the performance of data-driven models in alert-feedback systems. While oversampling techniques are commonly used to address this imbalance, they increase the total number of instances, leading to higher computational costs. Although the Active Learning (AL) approach is computationally expensive, it focuses only on the most informative samples, which can be more efficient for transactional records. Our experiments show that AL outperforms SMOTE and Borderline-SMOTE in terms of accuracy and AUPRC. Therefore, AL presents a promising approach for addressing the class imbalance problem in transactional records, without the added computational burden of synthetic samples.