Evaluating defect prediction approaches using a massive set of metrics: an empirical study