A Study of Rule Extraction from Double Random Forest to Identify the Characteristics of Working Poor in Jakarta Province, Indonesia

Abstract

Double Random Forest (DRF) outperforms Random Forest (RF) models, particularly when the RF model is underfitting. DRF generates more diverse and larger trees that significantly improve prediction accuracy. By applying association rule technique, the extracted rules from the DRF model provide an easily understandable interpretation of the characteristics of individuals identified as the working poor in Jakarta. The findings show that DRF performs good predictive performance in classifying poor workers in Jakarta, achieving an AUC value of 79.02%. The extracted rules from this model highlights interactions between education levels, working household member proportion, and job stability that significantly affect the classification of working poor. Specifically, worker with lower education levels, particularly high school or below, show a higher probability of being classified as poor workers. In addition, households with fewer employed members, especially those involving worker in self-employed/employee/freelancer roles, face a greater risk of falling into the poor category due to job instability and limited workforce participation. This implies that the interaction between the low proportion of working household members and low education, the interaction between unstable job position and low proportion of working household members, and the interaction between low education and unstable job position are the most important characteristics of the working poor in Jakarta.