Local Density-Based Adaptive Undersampling Approach for Handling Imbalanced and Overlapped Data

The issue of class imbalance poses a significant challenge to conventional machine learning. When processing imbalanced data sets, classical supervised learning algorithms often skew towards the majority class, causing inaccurate predictions for the minority class. The consequences of inaccurate prediction are severe when the information of the minority class becomes crucial. Moreover, when samples of varying classes with severe proportions overlap, the problem escalates. The majority of existing algorithms that focus on class imbalance and class overlap have demonstrated promising results but most of them are built upon the hypothesis that instances are uniformly distributed in the data space, ignoring the complexity of real-world data distribution. When real world data is incorporated, this deficiency hinders their capability to effectively identify overlapping regions, resulting in the loss of sample information. To address this concern, we incorporate density factors into classical methods for identifying overlapping regions, proposing the Local density-based adaptive undersampling approach for handling imbalanced and overlapped data (LDAU). LDAU first computes the local density for each instance, and then select the extent to which majority class instances should be removed for each overlapping region based on a comparison of the densities of the majority and minority classes. In doing so, it achieves clear boundaries for both the majority and minority classes while minimizing the degradation of majority class instance information.