<![CDATA[Abstract. Poverty is a major socioeconomic challenge in Indonesia that affects the effectiveness of social protection programs. In response to this challenge, the government has created social assistance programs to improve the welfare of the people. However, the distribution of social assistance is often considered to be inaccurate, resulting in households that are deemed eligible for social assistance not being identified as recipients. One solution to improve the accuracy of distribution is the application of machine learning in the context of classification. Several tree-based models, such as LightGBM, Random Forest, and XGBoost, were selected because of their superior capabilities compared to classical models such as logistic regression, especially in handling complex data and fulfilling model assumptions. This study compares the performance of these three models in predicting social assistance recipient status using data from the 2024 West Java Provincial National Socioeconomic Survey (SUSENAS). Model evaluation was conducted on several data pre-processing scenarios involving outlier handling, class balancing, and feature engineering. The results show that LightGBM consistently outperforms the other models on six metrics, namely Accuracy, Balanced Accuracy, F1-Score, ROC-AUC, PR-AUC, and Brier Score, out of a total of eight evaluation metrics used. SHAP analysis identifies Social Assistance History and Asset Score as the most influential features for model prediction. Friedman and Nemenyi nonparametric tests confirmed significant performance differences between LightGBM and other models based on the F1-Score, PR-AUC, and Brier Score metrics. These findings indicate that tree-based models, particularly LightGBM, can support the development of a more targeted and data-driven social assistance targeting system. Keywords: Social Assistance; Tree-Based; SHAP; SUSENAS; Hybrid Bayesian Optimization]]>
Copyrights © 2026
