Dynamic Boundary-Aware Incremental Ensemble Learner for Imbalanced Data Classification

Abstract

Classification of imbalanced data with high class overlap is a continuous problem in machine learning. Traditional methods bear either from noise addition through oversampling or information loss due to undersampling neither able to deal with the problems conjoined with overlapping class distributions. In this paper, we suggest a Dynamic Boundary-Aware Incremental Ensemble method that combines adaptive sampling, progressive learning, and dynamic ensemble weighting. The DBAIE works by a two-step dynamic boundary sampling approach that gives priority to high contribution samples close to the decision boundaries and progressive incremental learning using Elastic Weight Consolidation to evade catastrophic forgetting. Moreover, our adaptive ensemble framework further improve classification performance by dynamically accommodate classifier weights using real-time feedback. We guided extensive experiments on multiple datasets, which expose that DBAIE significantly surpasses state of the art methods in control complex class distributions and preserving high accuracy for minority classes. The DBAIE method presents a sturdy and scalable key for classifying imbalanced data with high precision, recall, and overall classification metrics.