Machine Learning Approaches for Household Smoking Classification: A Comparison between Encoded Logistic Regression and XGBoost

Abstract

Smoking behavior within households is a significant public health issue, particularly in densely populated regions such as West Java. This study aims to compare the classification performance of Logistic Regression and XGBoost in identifying household smoking status under several feature representation schemes, including baseline models, One-Hot Encoding, and Weight of Evidence (WoE). The analysis used 23.854 household observations from 2024 National Socioeconomic Survey (SUSENAS) in West Java after preprocessing. Model performance was evaluated using 5-fold stratified cross-validation based on accuracy, sensitivity, specificity, and Area Under the Curve (AUC) metrics. Among the evaluated configurations, XGBoost with WoE Encoding achived the highest observed accuracy (0.739), sensitivity (0.782), and AUC (0.797), although the differences in AUC across models were relatively small. Gain-based feature importance identified the Food–Non-Food Expenditure Ratio (X1) as the most influential predictor. The dominance of this expenditure ratio confirms that household expenditure structure, as an indicator of economic condition and welfare, plays a crucial role in distinguishing smoking households. These findings support the hypothesis that poorer households tend to have a higher probability of smoking.