Machine learning approach for multidimensional poverty estimation

In the social sciences, a theoretical analysis has predominated in its research. The scarcity of data and its difficulty in collecting and storing it, has been the main limitation for the social sciences to adopt quantitative approaches. However, the large amount of information generated in recent years, mainly through the use of the Internet, has allowed the social sciences to include more and more quantitative analysis. This study proposes the use of technologies such as Machine Learning (ML) are the answers to solving this data scarcity. The objective is to estimate the multidimensional poverty index at the personal level in a particular territory of Ecuador by using Machine Learning (ML) regression models based on a limited amount of data for training. Ten ML models are compared, such as linear, regularized, and assembled models and Random Forest performs outstandingly against the other models. An error of 7.5% was obtained in the cross-validation and 7.48% with the test data set. The estimates are compared with statistical approximations of the MPI in a geographical area and it is obtained that the average MPI estimated by the model compared to the average reported by the statistical studies differs by 1%.