Using Artificial Neural Networks to Visualize Poverty

From 69,130 households that were covered by a comprehensive community-based monitoring survey conducted in one of the cities that comprise Metro-Manila, in the Philippines, a neural network technique is used to identify the “absolute poor”. Households are considered to be among the “absolute poor” when the per capita income is less that 1USD per day, which is based on the UNESCO definition of absolute poverty. Based on this definition, 10% or 6,998 households are considered poor. A backpropagation neural network is trained to distinguish households as either poor or not. We achieve an accuracy of about 61% on both the train and test sets. Further rule-extraction on the trained network is done in order to understand, in terms of the features used for training, which features contribute to the positive identification of households that are poor by UNESCO definition. To complement the extracted rules, the poverty dataset is also used to train a Self-Organizing Map (SOM), which is then used to allow for an intuitive visualization of various facets of poverty. From the trained SOM, three distinct “poverty” clusters were identified.