An innovative hybrid approach for imbalanced datasets: IQCM methodology and comparative performance analysis

This study addresses the issue of unequal class distribution in real-world datasets, which affects the performance of classification algorithms. A new hybrid method (IQCM) is developed based on data isolation, clustering, distance correlation, and weighted arithmetic mean techniques to mitigate the issue of imbalance. This method consists of two primary steps. In the first step, the Isolation Forest algorithm is applied to the majority class during undersampling to detect outliers defined as potentially noisy samples. In the second step, Quantum Clustering is used to detect higher density regions of minority class samples. By applying the distance correlation to these samples, the closest pairs are determined in terms of similarity and relationship. In the final step, utilizing weighted arithmetic mean, a sufficient number of synthetic examples are generated between these pairs, resulting in the balancing of datasets. In the experimental phase, Turkish/English tweets about Covid-19 obtained from Twitter and numerical datasets with different imbalance ratios is used. Topic modeling and sentiment analysis are performed using the Non-Negative Matrix Factorization method in Twitter datasets. It is determined that the labeled Twitter datasets have an imbalanced situation. The proposed method and existing resampling methods (RUS, NearMiss, SMOTE, Borderline SMOTE, SMOTE-Tomek, SVM-SMOTE, KMeans-SMOTE) are tested with 2 textual and 20 numerical datasets. These datasets are balanced in the preprocessing phase and classified by Random Forest, Support Vector Machines and Logistic Regression algorithms. As a result of the study, methods that reduce the imbalance problem and increase the classification performance are determined and compared. The results indicated that the proposed method is significantly better than existing resampling methods in both textual and numerical data and can improve the prediction performance of classifiers. They yield the best in the average of F-measure, AUC-ROC and G-mean. According to these findings, it has been observed that the suggested hybrid approach is more effective in classifying imbalanced datasets.