SMOTE-IF: A Novel Resampling Method Based on SMOTE Using Isolation Forest Variants for Multi-Class Imbalanced Data
Ang Li, Tingting Ma, Sen Ye, Xunyun Liu · 2023
Imbalance learning is an important branch of classification tasks in the field of machine learning, and has received increasing attention from researchers. Currently, most researches have focused on binary imbalanced problems, while there exist numerous unsolved multi-class imbalanced problems in real world. The diversity of data distributions and the poor performance of traditional multi-class classification algorithms pose significant challenges to classify multi-class imbalanced data. In this paper, we propose a resampling method named SMOTE-IF based on isolated forest to address the issue of imbalanced overlapping in multi-classification tasks. Firstly, in order to reduce the negative impact of having severely few minority classes, we first propose a SMOTE-based strategy to oversample them. Secondly, a variant of isolated forest is proposed, which can identify overlapped and noisy data in multi-class of boundaries. Numerous experiments on various real datasets have shown that the SMOTE-IF method can effectively handle imbalanced overlapping data. Compared with state-of-the-art resampling methods, SMOTE-IF has achieved significant improvements in different classification performance metrics.