Targeting Social Assistance Beneficiaries Using Machine Learning: A Poverty Probability-Based Approach

Abstract

In a context where social inequalities are deepening and public resources are becoming increasingly scarce; the fair and effective identification of social assistance beneficiaries has become a central issue. Traditional targeting methods, such as categorical eligibility or proxy means testing, are now showing their limits, frequently producing inclusion and exclusion errors.

This study relies on a synthetic dataset of 12,600 individuals described by 59 socio-economic variables, ranging from demographic characteristics and education level access to financial and digital services. Three supervised learning models are compared: logistic regression, Random Forest, and XGBoost. The results reveal that tree-based models outperform logistic regression, particularly in reducing exclusion errors, which are especially critical in social policy contexts.

The analysis of key variables highlights the decisive role of education levels, place of residence (urban/rural), and access to digital and financial services. These findings confirm the need for a multidimensional approach to poverty that goes beyond purely monetary criteria. Finally, the study emphasizes the ethical challenges raised using algorithms: transparency, bias reduction, and institutional accountability emerge as essential conditions for legitimizing their integration into social protection and for contributing to more inclusive and equitable systems.

Classification JEL: I32; I38; C45; C55; H53

Paper type: Empirical Research