FuzzyCSampling: A Hybrid fuzzy c-means clustering sampling strategy for imbalanced datasets

Classification model with imbalanced datasets is recently one of the most researched areas in machine learning applications since they induce to the emergence of low-performing machine learning models. The imbalanced datasets occur if target variables have an uneven number of examples in a dataset. The most prevalent solutions to imbalanced datasets can be categorized as data preprocessing, ensemble techniques, and cost-sensitive learning. In this article, we propose a new hybrid approach for binary classification, named FuzzyCSampling, which aims to increase model performance by ensembling fuzzy c-means clustering and data sampling solutions. This article compares the proposed approaches' results not only to the base model built on an imbalanced dataset but also to the previously presented stateof-the-art solutions undersampling, SMOTE oversampling, and Borderline Smote Oversampling. The model evaluation metrics for the comparison are accuracy, roc_auc score, precision, recall and F1-score. We evaluated the success of the brand-new proposed method on three different datasets having different imbalanced ratios and for three different machine learning algorithms (k-nearest neighbors algorithm, support vector machines and random forest). According to the experiments, FuzzyCSampling is an effective way to improve the model performance in the case of imbalanced datasets.