Optimizing machine learning algorithms for multidimensional poverty prediction in the Philippines

This work aims to enhance poverty prediction and policy targeting using machine learning in the poorest region of Luzon, Philippines. Current poverty measurement methods are time-consuming, labor-intensive, and expensive. As a result, policy targeting becomes challenging for policymakers when implementing poverty alleviation programs. To address this, our study leverages community-based system datasets, applying machine learning regression and classification algorithms combined with advanced econometric models. For regression, we applied seven algorithms across 273 ensemble runs, while for classification, we employed twelve algorithms across 468 ensemble runs to analyze 34 localities and four sectors at a disaggregated level, incorporating cross-validation before combining results. Random forest regression outperformed all models, achieving an MSE of 0.0792, RMSE of 0.3298, and R2 of 0.92075. Similarly, the random forest classifier achieved the highest accuracy (91.08% in random selection and 95.95% in pipeline selection). Furthermore, our paper analyzes 27 multidimensional attributes in relation to key poverty indicators (incidence, gap, severity, and Watts index). This study demonstrates the feasibility of using machine learning for poverty prediction, offering a cost-effective, labor-efficient, and time-saving approach, particularly in the poorest regions of the Philippines. Finally, the study provides policy-targeting tools for poverty reduction across various localities with different poverty configurations.