pattern mining, commits, Machine learning on source code

Professional software developers write code in small chunks (called commits) which enable clean versioning of software. Best practices recommend that these chunks should be conceptually simple and should not mix many thematically different changes of software together. This contribution aims to present an analysis of change patterns in 6 million files which were extracted from 600 most popular software repositories written in Java. Every change pattern contains a set of edits in the file where every edit is represented by the type of change (insert, delete, move, update) and type of entity (for loop, a function call, if statement, etc.). We also include structural information of every edit (i.e. for loop was inserted into the body of a function). Based on these features, we run topic modeling to identify patterns of changes in files.