An Improved Over-sampling Algorithm based on iForest and SMOTE

Abstract

Imbalance learning is one of the most challenging problems in supervised learning, so many different strategies are designed to tackle balanced sample distribution. The over-sampling techniques which achieve a relatively balanced class distribution through synthesizing samples receive more and more attention. In this paper, we present an over-sampling approach based on isolation Forest (iForest) and SMOTE, called iForest-SMOTE. Firstly, for minority class samples, iForest-score is employed to assess the importance of each sample based on iForest model. Then, in each SMOTE process, roulette wheel selection based on iForest-score is utilized to select the neighbor sample. Finally, M-dimensional-sphere interpolation approach is employed to generate a new sample. The experiments illustrate that our approach takes into account the spatial distribution of minority class samples and sample synthetic simultaneously. Therefore, iForest-SMOTE can effectively improve the performance of the classification model.