Supervised learning algorithms for famine prediction

Abstract

Early detection of famine reduces the vulnerability of the society at risk. This study examined the application of supervised learning algorithms for famine prediction. Data were collected between 2004 and 2005 from households in northern, central, eastern, and southern parts of Uganda. Data sets from the northern region were the most suitable as a learning sample for other regions. Classification performance of Support Vector Machine, K-Nearest Neighbors, Naïve Bayes and Decision tree in prediction of famine were evaluated. Support Vector Machine and K-Nearest Neighbors performed better than the other methods, and Support Vector Machine produced the best Receiver Operating Characteristics (ROC), which can be used by policy makers to identify famine-prone households. It is recommended that satellite and household data should be used in combination to predict food security because this increases the specificity of households at risk. Copyright © 2011 Taylor & Francis Group, LLC.