Application of Random Forest with SMOTE and Random Search for Food Security Classification in Indonesia

Abstract

Food security is an important aspect of sustainable development and one of the key priorities of the Sustainable Development Goals (SDGs). Differences in food security conditions across regencies and cities in Indonesia require accurate classification to support the identification of vulnerable regions and appropriate
policy-making. The random forest method is a classification method with good predictive performance, which can be improved through hyperparameter optimization using random search. However, its classification performance may decline when applied to imbalanced datasets. Therefore, the Synthetic Minority Over-sampling Technique (SMOTE) was employed to address class imbalance. This study aimed to classify the food security levels of
regencies and cities in Indonesia in 2024 using random forest with SMOTE and random search hyperparameter optimization and to evaluate model performance based on accuracy, precision, recall, and F1-score. The data were obtained from the Food Security and Vulnerability Atlas (FSVA), comprising 514 regencies and cities, with food security level as the response variable and eight predictor variables. The dataset was divided into training and
testing sets using proportions of 80:20 and 90:10. SMOTE was applied with K = 5, while random search was conducted using 20 hyperparameter combinations to determine the optimal model. The results showed that the 80:20 data partition outperformed the 90:10 partition, achieving an accuracy of 88.35%, precision of 83.33%, recall of 50.00%, and F1-score of 62.50%. The best model correctly classified 10 food-insecure and 81 food-secure regencies and cities. These findings provide information for identifying regions based on food security levels and support more targeted food security policy-making.