Robust hybrid data-level sampling approach to handle imbalanced data during classification

Classification process is significant in finding different patterns from data. The performance of classifiers is highly affected with many data impurities like imbalance data, noise, class overlapping and different distributions of data within classes. The data in the real-world applications are often corrupted with multiple data impurities. To handle this issue, this paper proposed a hybrid data-level method to handle multiple data impurities like class imbalance, noise and different data distributions within classes. The proposed approach works in phases; in the first phase, it identifies and removes noise from the data, and then, it detects minority and majority cluster by using kernel-based fuzzy clustering approach. Radial basis kernel is used for clustering. In the next phase, minority and majority clusters are processed to balance the data. It uses radial basis kernel fuzzy membership and \(\alpha \)-cut to reduce the data size of majority cluster- and firefly-based SMOTE method to intelligently produce synthetic data within minority cluster. After removing all the data impurities, a traditional classifier (Decision Tree) is used to classify the balanced data. Performance of proposed method is tested with 3 synthetic data-sets and 44 UCI real-world data-sets of different imbalance ratios (imbalance ratio varies from 1.82 to 129.44). Area under the ROC curve is used to assess and compare the performance of proposed method with 20 other data-level methods. Experimental results confirmed that proposed method outperformed every other method especially in the case of highly imbalanced data-set.