How robust is a machine learning approach to improving food security amid the crisis? Evidence from Uganda's Covid-19
Cao, Gewei/Lukas Kornher/Clara Brandi
Externe Publicationen (2025)
In: World Development 196, Article 107171
doi: https://doi.org/10.1016/j.worlddev.2025.107171
Open Access
Amid a variety of global food insecurity challenges, including the Covid-19 pandemic and economic disruption, this article explores the potential of machine learning (ML) to enhance food instability predictions. So far, few existing studies have used pre-pre-shock training data to predict food instability, and if so, have not systematically tested the performance and robustness of the ML algorithm at the impact stage at the household level. To address this research gap, we propose a new approach to assess the performance and robustness of ML models using pre-Covid trained models to predict household-level food insecurity during the COVID-19 pandemic in Uganda. Therefore, the aim of this study is to find high performance and robust ML algorithms during shock periods. This is methodologically innovative and is indeed relevant to the study of food anxiety. First, we see that ML works well in the shock context if only food security data from Pre-Shock is available. To classify food-secured approximately 40% of food-secure households as unstable, we can identify 80% of food-affected households during the Covid-19 pandemic based on a pre-shock trained model. Second, we show that an extreme gradient boost algorithm trained with balanced weighting works best in terms of prediction quality. They also identified the most important predictors and found that demographics and asset characteristics play an important role in predicting food instability. Finally, we also contribute to various ML models by showing how different ML models should be evaluated in terms of curve (AUC) values (AUC) values, the ability to correctly classify positive and negative cases, and the ability to evaluate different ML models in terms of AUC changes in different situations.
