Application of SMOTE on CART Method to Handle Imbalanced Data (Study Case: Labor Force Classification in Banten Province)

Abstract

Publication of the Central Bureau of Statistics (BPS) show that Banten is province with the highest unemployment rate in Java Island during the period 2006 to 2016. One of the efforts in resolving this issue is to do a classification of labor force into unemployment and employment as well as identify its characteristics, so that later the government policy will not be mistaken. The classification method used in this research is CART that has the ease of interpretation against the results of the analysis. The accuracy of a classification tree can be viewed from the value of specificity, sensitivity, and the AUC. Unfortunately, the imbalanced data causes low value of sensitivity and AUC. One alternative way to increase the accuracy is to conduct SMOTE on pre processing data. Classification tree with SMOTE is more accurate compared to the classification tree without SMOTE, because it can increase sensitivity from 6.35% to 77.78% and has higher AUC, as much as 0.7846. While AUC of classification tree without SMOTE is 0.7507. The best classification tree produce 18 terminal knots that influenced by 6 variables, which are age, gender, marital status, status in the household, level of education completed, and the residence area.