A Machine Learning Approach to Determine the Semantic Versioning Type of NPM Packages Releases

Semantic versioning policy is widely used to indicate the level of changes in a package release. Unfortunately, there are many cases where developers do not respect the semantic versioning policy, leading to the breakage of dependent applications. To reduce such cases, we proposed using ML techniques to effectively predict the new release type, i.e., patch, minor, major, in order to properly determine the semantic versioning type. To perform our prediction, we mined and used a number of features about a release, such as the complexity of the changed code, change types, and development activities. To evaluate the performance of the proposed ML classifiers, we conducted an empirical study on 31 packages containing a total of approximately 6,260 releases. We extracted 41 release-level features from historical data of packages’ source code and repositories. We used four ML classifiers. We found that the XGBoost classifiers performed the best. We also found that features related to the change types in a release are the best predictors group of features in determining the semantic versioning type. Finally, we studied the generalizability of determining the semantic versioning type by applying cross-package validation. Our results showed that the general classifier achieved a good performance.