Usage-based Models Improve Odds in Software Testing