AI for Good

Chapter 17Food Security Analysis

—Shahrzad Gholami, Erwin Knippenberg, and James Campbell

Executive Summary

Exacerbated by climate change–driven shocks like droughts and floods, food insecurity—the lack of consistent access to enough food for every person in a household to live an active, healthy life—is a global problem. For humanitarian programs to ensure timely delivery of assistance, forecasting food insecurity levels and identifying vulnerable households is crucial. Here, we used a machine learning approach trained on high-frequency household survey data to identify predictors of food insecurity at the community level and forecast household-level risks for food insecurity in near real-time.

To develop predictive models, we used data collected monthly through the Measuring Indicators for Resilience Analysis data collection protocol implemented by Catholic Relief Services in southern Malawi. We considered the predictive model as a binary classification of food insecurity that we dichotomized based on two different thresholds, which resulted in two different positive class to negative class ratios (with one class being food insecure and the other class being not food insecure).

When predicting community-level vulnerability, we found that a random forest model outperformed other approaches and that location and self-reported welfare are the best predictors of food insecurity; the model was accurate in predicting food insecurity when predictor features included a historical ...