Improving clustering-based and adaptive position-aware interpolation oversampling for imbalanced data classification

Class imbalance is one of the most significant difficulties in modern machine learning. This is because of the inherent bias of standard classifiers toward favoring majority instances while often ignoring minority instances. Interpolation-based oversampling techniques are among the most popular solutions for generating synthetic minority samples to correct imbalanced class distributions. However, synthetic minority samples have a risk of overlapping with the majority-class samples. Inappropriate interpolation of minority samples during oversampling can also result in over generalization. To overcome these drawbacks, we propose a Clustering-based and Adaptive Position-aware Interpolation Oversampling algorithm (CAPAIO) for imbalanced binary dataset classification. CAPAIO initially employs an improved density-based clustering algorithm to group minority instances into inland, borderline, and trapped samples. It then adaptively determines the size of each subcluster and allocates weights to minority samples, guiding the synthesis of minority samples based on these weights. Finally, distinct interpolation oversampling algorithms are individually performed on these three categories of minority samples. The experimental results demonstrate the effectiveness of the proposed CAPAIO in most datasets compared with eleven other oversampling algorithms.