An empirical study on reducing omission errors in practice

Abstract

Since studies based on mining software repositories sparked interests in the field of guiding software changes, many change recommendation techniques have been proposed to reduce omission errors. While these techniques only used existing software commit data sets to evaluate their effectiveness, we use the data set of supplementary patches which correct initial incomplete patches to investigate how much actual omission errors could be prevented in practice. We find that while a single trait is inadequate, combining multiple traits is limited as well for predicting supplementary change locations. Neither does a boosting approach improve accuracy significantly, nor filtering based on developer or package specific information necessarily improves the accuracy. Developers rarely repeat the same mistakes, making the potential value of history-based change prediction less promising. We share our skepticism that omission errors are hard to prevent in practice based on a systematic evaluation of a supplementary patch data set.