A Pruning-Based Approach for Searching Precise and Generalized Region for Synthetic Minority Over-Sampling

Abstract

One solution to deal with class imbalance is to modify its class distribution. Synthetic over-sampling is a well-known method to modify class distribution by generating new synthetic minority data. Synthetic Minority Over-sampling TEchnique (SMOTE) is a state-of-the-art synthetic over-sampling algorithm that generates new synthetic data along the line between the minority data and their selected nearest neighbors. Advantages of SMOTE is to have decision regions larger and less specific to original data. However, its drawback is the over-generalization problem where synthetic data is generated into majority class region. Over-generalization leads to misclassify non-minority class region into minority class. To overcome the over-generalization problem, we propose an algorithm, called TRIM, to search for precise minority region while maintaining its generalization. TRIM iteratively filters out irrelevant majority data from the precise minority region. Output of the algorithm is the multiple set of seed minority data, and each individual set will be used for generating new synthetic data. Compared with state-of-the-art over-sampling algorithms, experimental results show significant performance improvement in terms of F-measure and AUC. This suggests over-generalization has a significant impact on the performance of the synthetic over-sampling method.