Abstract

In a manual examination of more than 7,000 issue reports from the bug databases of five open-source projects, we found 33.8% of all issue reports to be misclassified, that is, rather than referring to a code fix, they resulted in a new feature, an update to documentation, or an internal refactoring. This misclassification introduces bias in bug prediction models, confusing bugs and features: On average, 39% of files marked as defective actually never had a bug. We estimate the impact of this misclassification on earlier studies and recommend manual data validation for future studies.

‚Äč