Prediction of Multidimensional Poverty Status With Machine Learning Classification at Household Level: Empirical Evidence From Tanzania

Over fifty percent of the population in Tanzania suffers from multidimensional poverty. Because of the high poverty rate and slow improvement, ending poverty by the year 2030 remains challenging and empirically testable proposition and part of a shared challenge. The main purpose of this study is to predict multidimensional poverty status with the help of best performance-supervised machine-learning algorithms. To achieve this objective, longitudinal data from the 2014/15 and 2020/21 surveys, sourced from the Tanzania National Bureau of Statistics (NBS) were analyzed. A variety of supervised machine-learning algorithms such as RBF Kernel in SVM, Linear Kernel in SVM, Polynomial Kernel in SVM, Random Forest, Logistic regression classifier, Decision tree, Gradient Boosting, K-Nearest Neighbours Classifier, Naïve Bayes Classifier, Artificial Neuron Network and Ensemble Learning Model were implemented to predict multidimensional poverty status for each dataset. This captured the dynamic changes from 2014 to 2021 at the national level. The study employed data pre-processing techniques and adjusted imbalance through weighted categorical cross entropy and 5-Fold Cross Validation to mitigate the inefficiencies due to overfitting and under-fitting of the algorithms. Additionally, dimensionality data reduction was performed using Principal Component Analysis (PCA). With regard to the evaluation metrics, we show that the Ensemble Learning Model has achieved the best performance modelling both balanced and unbalanced datasets. Our policy recommendations draw on the results of the algorithmic predictions.