Machine Learning-Based Poverty Risk Analysis for Targeted Government Financial Assistance

Machine learning has been widely applied to poverty risk analysis at national and regional levels, but its use at the barangay level remains limited, particularly in the Philippine context, where data are collected by Barangay Health Workers (BHWs). This study addresses this gap by analyzing data from 973 households in Barangay Cogon, El Salvador City, aiming to identify key poverty determinants and evaluate the predictive performance of Decision Tree (DT), Random Forest (RF), and Artificial Neural Network (ANN) models for poverty risk classification. Results show that monthly household income, family size, and parental education are the most significant predictors of poverty risk. Across all evaluation settings, RF consistently outperformed DT and ANN, achieving the highest values in all performance metrics. Under an 80% training split, RF achieved an AUC of 0.839, accuracy of 0.793, precision of 0.809, F1-score of 0.800, and recall of 0.791. With a 70% split, RF obtained an AUC of 0.827, accuracy of 0.790, precision of 0.817, F1-score of 0.793, and recall of 0.771. In five-fold cross-validation, RF maintained strong performance with an AUC of 0.833, accuracy of 0.791, precision of 0.807, F1-score of 0.792, and recall of 0.778. These findings indicate that the Random Forest model is a robust and effective decision-support tool for poverty risk classification and is recommended for adoption in local targeting systems to improve the accuracy of identifying beneficiary households. Furthermore, future studies are also recommended to enhance model performance by incorporating additional multidimensional indicators such as health status, access to social protection programs, and housing quality to provide a more comprehensive assessment of poverty risk.