A Gaussian Mixture Model–Driven Hybrid Resampling Technique for Imbalanced Data Classification

Abstract

The issue of class imbalance is prevalent in various real-world applications, including spam filtering, fraud detection, and medical diagnosis. It has a significant impact on how well learning algorithms function. Under sampling at random is a well-known way to overcome the issue. However, because of their randomness, it is unable to effectively extract the samples close to the majority and minority class cross-edge, despite the fact that these samples are crucial to a classifier since they affect the classification performance. This class rarity leads to classification methods behavior biased towards majority classes that cause misclassification. There are different methods for handling imbalanced data like data-level, algorithm-level, and hybrid. For handling this issue we propose a GMM-based approach, which is performed in three ways as GMM- based under sampling, GMM-based oversampling, GMM-based combination of under sampling and oversampling. For evaluation of results we use four matrices such as precision, recall, f-measure, and accuracy after GMM-based oversampling precision, recall, F1 is 83%, 86%, 84% in fraud detection respectively, accuracy is 0.43 and in GMM-based under sampling precision, recall, F1is 0%, 100% and 1% and the accuracy is 0.50, while combining the GMM-based under sampling and oversampling the precision, recall, F1is 0%, 99%, 1% respectively and the accuracy of fraud detection is 0.43.