The incremental SMOTE: A new approach based on the incremental k-means algorithm for solving imbalanced data set problem

Abstract

Classification is one of the very important areas in data mining. In real-life problems, developed methods for modeling with the classification problem generally perform well on datasets where the class distribution is balanced. On the other hand, the data sets are often imbalanced and it is important to develop algorithms to solve the classification problem on imbalanced data sets. Imbalanced datasets are more difficult to classify than balanced datasets because learning a class with underrepresentation is difficult. Most real life problems are imbalanced. The class with the least number of data usually corresponds to rare cases and is more important. Learning these classes is critical accordingly. One of the most commonly used solution methods to solve this problem is to oversample the minor class. When oversampling, too many repetitions in the dataset can cause overfitting. For this reason, it is very important to ensure data diversity when oversampling. Therefore, this paper proposes a new oversampling methods (the incremental SMOTE) combining the incremental k-means algorithm and Synthetic minority oversampling technique (SMOTE). The original dataset is clustered with the incremental k-means algorithm and the clusters are filtered to determine the safe clusters. The number of points to be produced from the safe clusters is determined, and then new instances are produced with the improved SMOTE algorithm. In the incremental SMOTE, diversity in the dataset is achieved by generating with incremental rate. In order to evaluate the performance of the incremental SMOTE algorithm, classification was performed on imbalanced datasets, balanced datasets obtained by the random oversampling, SMOTE, Borderline-SMOTE and SVM SMOTE methods. Comparisons for 10 datasets showed that the performance of the proposed method improves as the imbalance ratio of the dataset increases.