Researchers reveal bias in a widely used measure of algorithm performance