Improving Imbalanced Data Classification Through Integrated Data-Level and Algorithm-Level Techniques

Imbalanced class distribution is common issue in machine learning and data mining. It affects various applications like fraud detection, medical diagnosis, and network intrusion detection. The mentioned problem occurs when the majority class greatly outnumbers the minority class, hindering learning systems designed for balanced distribution. Furthermore, presence of overlapping instances between two classes exacerbates problem. This paper introduces a novel method that integrates ensemble and cost-sensitive techniques at algorithm-level with OSBNR approach at data-level to address this challenge. Furtherly, suggested technique is compared with conventional resampling methods, in conjunction with three ensemble algorithms and three cost-sensitive algorithms, to determine optimal combination. Simulation presents OSBNR is having optimal results in terms of processing time and evaluation metrics like G-mean, AUC-ROC, TPR-Recall, precision, Fl-score and FPR. PyOSBNR is simulated as python index package which can easily handle class imbalance in the presence of behavioral noise.