Implementation of Winsorizing and random oversampling on data containing outliers and unbalanced data with the random forest classification method

Abstract

Many researchers conduct research using the classification method, to find out the best method for predicting the class of an observation. Some of these studies explain that random forest is the best method. However, the classification of data containing outliers and unbalanced data is a complicated problem. Many researchers are also conducting research to deal with these problems. In this study, we propose a winsorizing to deal with outliers by replacing the outlier values with the upper and lower limit values obtained from the interquartile range method and random oversampling to balance the data. It is also known that cases of the Human Development Index (HDI) in regencies/cities in eastern Indonesia vary widely, so cases of HDI in these areas can be used as case studies of data containing outliers and unbalanced data. The purpose of this study was to compare the performance of the random forest before and after the data were applied to the winsorizing and random oversampling to predict HDI in districts/cities in eastern Indonesia. Classification method random forest after handling data containing outliers and unbalanced data has better performance in terms of accuracy and kappa values, which are 96.43% and 93.41%, respectively. The variables of expenditure per capita and the mean years of schooling are the most important.