Abstract: The contemporary real-world datasets often suffer from the problem of class imbalance as well as high dimensionality. For combating class imbalance, data resampling is a commonly used approach whereas for tackling high dimensionality feature selection is used. The aforesaid problems have been studied extensively as independent problems in the literature but the possible synergy between them is still not clear. This paper studies the effects of addressing both the issues in conjunction by using a combination of resampling and feature selection techniques on binary-class imbalance classification. In particular, the primary goal of this study is to prioritize the sequence or pipeline of using these techniques and to analyze the performance of the two opposite pipelines that apply feature selection before or after resampling techniques i.e., F + S or S + F. For this, a comprehensive empirical study is carried out by conducting a total of 34,560 tests on 30 publicly available datasets using a combination of 12 resampling techniques for class imbalance and 12 feature selection methods, evaluating the performance on 4 different classifiers. Through the experiments we conclude that there is no specific pipeline that proves better than the other and both the pipelines should be considered for obtaining the best classification results on high dimensional imbalanced data. Additionally, while using Decision Tree (DT) or Random Forest (RF) as base learner the predominance of S + F over F + S is observed whereas in case of Support Vector Machine (SVM) and Logistic Regression (LR), F + S outperforms S + F in most cases. According to the mean ranking obtained from Friedman test the best combination of resampling and feature selection techniques for DT, SVM, LR and RF are SMOTE + RFE (Synthetic Minority Oversampling Technique and Recursive Feature Elimination), Least Absolute Shrinkage and Selection Operator (LASSO) + SMOTE, SMOTE + Embedded feature selection using RF and SMOTE + RFE respectively.
