Machine Learning-Based System to Predict Poverty - Stricken Households in Rural Areas: The Case of Zimbabwe

In most countries in the world, poverty eradication is the most common subject, but what differs is the methods and techniques applied in identifying those who are in poverty. Different techniques which include night satellite images, surveys and censuses are among the methods used to try and identify population who are in different categories of poverty such as Extremely poor, Moderate poor and poor. Levels of poverty cannot be measured using the same indicators. Different indicators and different methodologies are applied across different countries in an effort to identify and classify households in poverty. The poverty measuring indicators vary depending on whether the household is in rural or urban area. The use of obsolete methodologies and data collection and processing technology lead to delays and error prone in the identification and delivery of government assistance to the households who are in extreme poverty. This study aims at developing a machine learning web-based system for classifying poverty-stricken households in rural areas of Zimbabwe. Among the supervised machine learning algorithms that was considered, Logistic Regression algorithm was the best poverty-stricken household classifier. Household Targeting Surveys (HTS) that was done by Ministry of Social Welfare and Zimstat since 2011 up to 2022 was the source of dataset used to train the model. For the purpose of identifying and classifying poverty-stricken rural households in Zimbabwe, indicators such as assets ownership, land ownership, household size, health status of head of household etc. was considered for the classification of households. See Fig. 1 for complete list of features that was extracted and considered. After train Logistic Regression algorithm on the HTS dataset, the model performance was 99% and this was achieved after performing model evaluation techniques such as cross validation.