<![CDATA[The need for data-driven analysis to understand socio-economic vulnerability, particularly in relation to food security, has intensified due to the pandemic. The growing volume of survey data demands analytical methods that can capture multidimensional relationships. Over time, food security evolves into the concept of food resilience, reflecting a household's capacity to withstand or recover from adverse conditions. This study uses machine learning techniques to categorize household food resilience, based on data from the World Bank's High Frequency Phone Survey (HFPS), covered 2,868 households in Indonesia. A comparative evaluation of gradient boosting algorithms (XGBoost, LightGBM and CatBoost) was conducted. Model performance was evaluated using accuracy, sensitivity, specificity, and AUC across ten repeated train-test splits, with statistical significance assessed using Friedman and Wilcoxon tests. The results show that CatBoost performed best and most consistently, achieving mean accuracy of 0.7592 and mean AUC of 0.8331, which is significantly higher than that of competing models. SHAP analysis further indicates that baseline vulnerability, financial concerns, income capacity, food price shocks and unmet healthcare needs are important features for identifying resilient households. These findings demonstrate that gradient boosting, particularly CatBoost, provides strong predictive power and interpretability to support data-driven decision-making in classifying food resilience.]]>
