Uma análise preditiva da insegurança alimentar brasileira utilizando um modelo de machine learning

Abstract

This research analyzed the risk profiles associated with food insecurity for Brazilian households, using microdata from PNADC 2024 and the modeling of a machine learning model XGBoost. The Logit econometric model was used for comparative purposes, allowing to examine differences in performance between inferential and predictive approaches. The analysis considers factors associated with household income, family composition, characteristics of the head of household, housing conditions, geographical location and receipt of social benefits. The results indicate that the XGBoost model presents better predictive performance than Logit, with greater accuracy and better balance between sensitivity and specificity. Household income per capita is highlighted as the main predictor of food insecurity, followed by receipt of social benefits, regional characteristics, color and education of the head of household and adults, occupational insertion and family composition. The study contributes to the deepening of the debate on food insecurity in Brazil by incorporating predictive modeling techniques still little explored in the national literature, offering empirical evidence that can subsidize the improvement of public policies aimed at mitigating food vulnerabilities.