Poverty level characterization via feature selection and machine learning

Abstract

A persistent socio-cultural problem of mankind is 'poverty', which requires accurate characterization in order to construct well designed policies for intervention. Unfortunately, the categorization along the poverty - wealthiness scale is not simply determined by applying surveys. Population is large, subjective opinions are usually biased, and available data are only indirectly related. In this paper, we attempt to identify poverty levels using feature selections from these indirect observations and machine learning techniques. In poverty assessment, similar to many other classification problems, it is crucial to know how any feature contributes to the classification of each class of poverty. We designed an approach that (1) extracts a subset of features that best characterize each poverty class, (2) examines how this subset affect the chosen class and finally (3) employ ensemble models. In this research, we adopt the Proxy Means Test (PMT) for labeling the data that was obtained from the Inter-American Development Bank of Costa Rica. Through this approach we analyze poverty classes within a multidimensional feature space perspective, contrary to the classically used single dimensional perspective defined as 'living on a consumption expenditure of less than the predefined income threshold'. The application and usefulness of our proposed framework is tested on the mentioned dataset using 85-15 data folding