Researchers use machine learning on household surveys to optimize global antipoverty program targeting and costs