Improving interpolation-based oversampling for imbalanced data learning

Abstract

The class imbalance problem in machine learning contains a skewed distribution of samples among different classes, resulting in a learning bias toward the majority class. Interpolation-based oversampling techniques solve this problem by yielding the synthetic minority samples to correct the imbalanced class distribution, which has become one of the most popular types of solutions. However, several non-negligible drawbacks exist in them. This study identifies that the flaws of the over constraint, low-efficiency expansion, and over generalization can occur when interpolating the synthetic samples for the inland minority samples (resided in the interior of dense regions of minority class), borderline minority samples (settled in the border of dense regions), and trapped minority samples (located in the sparse regions), respectively. To overcome these flaws, a Position characteristic-Aware Interpolation Oversampling algorithm (PAIO) is proposed. PAIO first leverages a neighborhood-based clustering algorithm to divide the minority samples into three types of samples with different position characteristics (i.e., the inland minority, borderline minority, and trapped minority samples). It then carries out ad hoc interpolation oversampling for these three different types of minority samples, so that the deficiencies can be addressed correspondingly. Moreover, another drawback in interpolation-base oversampling algorithms is the groundless populating for the categorical attributes of synthetic samples. To solve this problem, we develop a Generalized Interpolation Creation way (GIC) to fill the categorical attributes. Extensive experiments on a number of real-world datasets demonstrate that the effectiveness of the proposed PAIO and GIC.