Enhancing poverty classification in developing countries through machine learning: a case study of household consumption prediction in Rwanda

To address the challenges associated with measuring and classifying household consumption (poverty) in developing countries, such as cost, time gaps, and inaccurate socio-economic data, this study suggests leveraging machine learning (ML) algorithms. We assessed the performance of various ML algorithms using data from 14,580 sample households from the Integrated Household Living Condition Survey (EICV5), considering 87 features. Among the 12 classifiers evaluated, multiple kernel support vector machines, eXtreme gradient boosting, and multinomial logit demonstrated the highest predictive accuracy, ranging between 86.6% and 88.5%. Notably, household food expenditure, the total number of children (<14 years) in the household, and household own food expenditures emerged as the most predictive features for consumption classification. Interestingly, including shock-coping strategies did not significantly improve prediction accuracy. The multiple kernel support vector machine consistently outperformed eXtreme gradient boosting and multinomial logit. These findings suggest that survey questions used to assess poverty in Rwanda could be streamlined, prioritizing important features, particularly those related to household food characteristics. This approach has the potential to address challenges associated with measuring and classifying household consumption in developing countries more effectively.