This extended proposal assessed the performance of various machine learning algorithms for better prediction of household consumption (low-consumption, high-consumption but insecure, and high-consumption) in Rwanda using 14,580 sample households from 1,260 sample villages from a recent integrated household living condition survey (EICV5), various open-source MLAs were compared considering wide-ranging features (87). We empirically evaluated the 12 classifiers. The main findings are as follows: (a) household food expenditure, the total number of children (< 14 years) at the household level, and household own food expenditures are the most predictive features for household consumption; (b) multiple kernel support vector machine, eXtreme gradient boosting, and multinomial logit have significantly higher predictive accuracy between 86.6 % and 88.5 %; and (c) the inclusion of shock-coping strategies does not necessarily improve prediction. Therefore, some survey questions used to assess poverty worldwide could be reduced and prioritized for important features, such as household food characteristics.

Accepted Oral Presentation