Combining Unsupervised and Supervised Learning to Predict Poverty Households in Sakon Nakhon, Thailand

Abstract

Poverty is a problem that various government agencies are attempting to address accurately and precisely. This solution relies on data and analysis of features affecting poverty. Machine Learning is a technique to analyze and focus on poverty features encompassing five livelihood capitals: human, physical, economic, natural, and social capital to understand the household context and environment. The dataset contains 1,598 poverty households from Kut Bak district, Sakon Nakhon, Thailand. K-prototype was used to group categorical and numerical dataset into four clusters and labelled as Destitute, Extreme poor, Moderate poor, and Vulnerable non-poor. The performances of the Decision tree classifier with feature selection algorithms, including MI, ReliefF, RFE, and SFS, are compared. The best performance is SFS with F-measure, precision, and recall at 74.6%, 74.8%, and 74.7%, respectively. The result is the decision tree rules to predict the poverty level of households, enabling the establishment of guidelines for resolving household issues, and addressing broader problems within the areas.