Tomek link Undersampling with Stacked Ensemble classifier for Imbalanced data classification

Abstract

Imbalanced data is referred as unequal distribution of data between classes as a result of classifier. Majority class have higher percentage of data than minority class .Predicting the classifier accuracy is mostly biased toward the majority class, this leads to class imbalance problem. Misprediction of minority class has great loss in financial sector and life threatening problem in medical field. To solve the issue imbalanced dataset should be balanced by means of oversampling the minority class data or  undersampling the majority class data. Oversampling can cause over fitting of data. Among many undersampling algorithm Tomek link under sampling algorithm remove the noisy and borderline samples from majority class without losing important information of data. Tomek link undersampling algorithm can be applied to  4 benchmark Imbalanced dataset from UCI repository such as breast-W, breast cancer, hepatitis, heart diseases and so on. Stacking Ensemble classifier combine the output of individual classifier and fed to next level meta classifier to predict the result. First level prediction has been done using individual classifier such as NaiveBayes, Logistic Regression etc the output of first level fed to Meta classifier which predict the final class label. Performance measures such accuracy, precision, recall and auc_roc score are measured and compared with state of art classifier such as Support vector Machine, Decision tree, Naïve Bayes Classifier, K-Nearest Neighbour .Tomek link undersampling data applied to stacking classifier predict accuracy of 0.94 and roc_auc score of 0.97 which outperform the individual classifier performance measure