Selection of the Best Machine Learning Model to Predict Poverty Conditions: A Study on North-Eastern Wetland Region of Bangladesh

Abstract

Machine learning (ML) algorithms are effective techniques for predicting households’ poverty conditions so that they might benefit from poverty alleviation programs. The study’s primary objective is to find out the determinants of poverty and select the best ML model to predict the poverty conditions of the north-eastern wetland region of Bangladesh. This study used data from 2340 households that were collected through a household survey by a research project sponsored by the GARE Program, Ministry of Education, GoB. The multiple logistic regression (MLR) model was employed to extract the factors associated with household poverty. Six ML algorithms, including support vector machine, Naïve Bayes, logistic regression, K-nearest neighbor, decision tree, and random forest were applied to predict poverty conditions, and their performances were measured by using accuracy, precision, recall, F1-score, and AUROC. The study’s findings show that district, micro-credit status, household size, age, NGO membership, marital status, per capita income, cultivable land, electricity connection, and livestock ownership are the significant determinants of wetland people’s poverty. The findings also show that the support vector machine is the best model for predicting poverty level LPL with an accuracy of 82%, F1-score of 59%, and AUROC of 72%, and the logistic regression is the best model for predicting poverty level UPL with an accuracy of 81%, F1-score of 84%, and AUROC of 80%. The proposed algorithms may help improve poverty conditions by accurately predicting target poor groups. The determinants may be effective in developing policies to lessen poverty in the wetland region of Bangladesh.