Comparative Study of Predictive Classification Models on Data with Severely Imbalanced Predictors

Abstract

Analysing pre-COVID-19 unemployment in West Java is vital for comprehending and tackling Indonesia’s economic challenges. This significance arises not only due to the region’s high unemployment rate, but also from the need to understand unemployment patterns before COVID-19, which has become more relevant now during the country’s post-pandemic recovery phase. This study evaluates four machine learning models (Random Forest, Linear SVM, RBF SVM, and Polynomial SVM) to classify employment status using demographic and job-related variables. The objective is to find the most suitable model, particularly considering the imbalanced nature of the study-case data. Data from the National Labor Force Survey (SAKERNAS) in August 2019 is utilized, comprising 54,429 respondents across districts in West Java. The four models are evaluated using holdout validation with a 70:30 stratified proportion, repeated for 100 times. Results indicate that the random forest model outperforms others in balanced accuracy, F1-score, and computational time. The random forest model also underscores the importance of gender and age in classifying employment status in West Java, suggesting a need for targeted intervention, especially for female citizens and individuals in productive age groups.