Abstract Food insecurity is a frequently occurring country-wide problem caused by shocks with adverse outcomes on health and the economy. The objective is to predict data-driven levels of household food insecurity by considering the effect of space, time, and statistically significant factors. This study analyzed Ethiopian socio-economic household panel data for the years 2012, 2014, and 2016, comprising a sample of 3,835 households. Predictions were made using transductive deep learning algorithms to capture spatial and temporal dynamics, with 75% of the data allocated for training and 25% for testing. The results from training the temporal graph neural network model (T-GCN) with 4- and 2-layer architectures using 64 and 32 hidden units indicated that the 4-layer T-GCN with 64 hidden units predicted household food insecurity levels with greater accuracy. The prediction of the test data using the T-GCN model with 64 hidden units and 4 layers (T-GCN(64,4)) achieved an accuracy of 75.81%, outperforming baseline models in capturing spatial dependency with GCN (69.72%) and temporal correlation with RNN (54.66%). Additionally, the T-GCN(64,4) effectively predicted household food insecurity across most parts of the country, with an accuracy of 79.77% for 2014 and 72.43% for 2016. The higher stable prediction accuracy of the T-GCN in all prediction horizons demonstrated its superior performance and indicated the effectiveness of the data-driven approach in leveling severity and determining the related factors of household food insecurity. Therefore, efforts should be focused on urbanization, education, employment, non-agricultural business, water and sanitation, soil protection, and farming practices to reduce the severity of food insecurity by addressing dependency ratios, shocks, land ownership, and early marriage.
Wubetie et al. (Mon,) studied this question.