How a fraud algorithm learned to suspect vulnerable groups