Combining open data and machine learning to predict food security in Ethiopia

Food security is commonly measured by means of surveys, requiring substantial time and budget. Open data can possibly serve as a cost-effective alternative to predict food security. In this paper a method is proposed that uses open data related to food insecurity drivers to predict food security in Ethiopia at the subnational level. The method is based on an ordinal classification approach with a random forest as underlying algorithm. The model turned out to have an accuracy of approximately 90%. Although using an ordinal approach increases performance, a negative side-effect is that the model struggled to predict records with the label ‘stressed’ as a target. The basis of this effect lays in how probabilities for classes ranked in the middle are calculated. Further research on adding open data sources on other drivers and on finetuning hyperparameters in the modelling is advised before implementing machine learning to predict food security.