The classification of imbalanced datasets has garnered a significant research attention over the past decades. Although, numerous solutions have been proposed to address imbalanced classification problems, recent research suggests that class imbalance alone does not significantly degrade learning performance. Rather, class overlap, which has often been overlooked in previous studies has an even greater negative impact on the performance of learning algorithms. This paper proposes a Boosting-Inspired Overlap Reduction (BiOR) algorithm to handle the problem of class overlap in imbalanced datasets. BiOR employs a variance-based adaptive threshold mechanism, where the overlap threshold is dynamically determined as a function of the variance of margins in both majority and minority classes. By iteratively adjusting weights of samples, BiOR  ensures selective removal of highly overlapping majority instances, thereby enhancing class separability without excessive data loss. The effectiveness of BiOR is evaluated on 20 benchmark real world datasets as well as 16 synthetic datasets, demonstrating its superior classification performance over existing resampling techniques.