EVALUATION OF MULTIVARIATE ADAPTIVE REGRESSION SPLINES ON IMBALANCED DATASET FOR POVERTY CLASSIFICATION IN BENGKULU PROVINCE

Abstract

Classification is a statistical method that aims to predict the class of an object whose class label is unknown. The Multivariate Adaptive Regression Splines (MARS) classification method is a classification model that involves several basis functions with influential predictor variables. The MARS classification model is generally effective in classifying imbalanced data, including poverty data classification. The response variable used is the poverty status of households classified into poor and non-poor households, and the predictor variables consist of several poverty indicators. The problem that often arises in classification methods is a class imbalance in the response variable. Due to the poverty status included in the class imbalance data, the Bootstrap Aggregating (Bagging) and Synthetic Minority Over-sampling Technique (SMOTE) approaches will be used to improve classification accuracy on the MARS model. Bagging works by replicating data to strengthen the stability of classification accuracy, while SMOTE works by synthesizing data from minority data classes. The evaluation results showed that the classification model of poverty in Bengkulu Province using the SMOTE-MARS method provides the best classification accuracy compared to the MARS (25.81%) and Bagging-MARS (32.26%) methods based on the sensitivity value obtained, which is 85.36%.