Usage of Embedding-based Topic Modeling for Feature Extraction in Software Engineering Tasks

The practical application of machine learning models for tasks in the Software Engineering domain requires high models’ accuracy. However, according to the existing work, even high accuracy does not guarantee practical usefulness when the predictions are shown to real developers. One possible way to solve this problem is the usage of interpretable models. A common approach to extracting information in interpretable form from large corpora of textual data is topic modeling. Previous work on topic modeling in the Software Engineering domain employed topic modeling algorithms directly transferred from Natural Language Processing field. This work introduces Code2Topic, an algorithm for topic modeling of source code that takes into account characteristics typical to code such as large vocabulary size. Code2Topic can transform any piece of code into a distribution of topics in it. This information can be further used in practical tasks. Based on the Code2Topic algorithm, this work presents Sosed, a tool for finding similar repositories among a database of 9 million open-source projects. Sosed transforms projects into topic distributions and searches for similar distributions treating them as vectors. Manual labeling of Sosed’s predictions for a dataset of 94 popular GitHub projects proved that its predictions are relevant: average relevance among top-5 predictions is 4.2 out of 5. Another contribution of this work is Dev2Topic, an adaptation of the Code2Topic algorithm to work with the code of individual developers. It allows representing developers’ expertise as a distribution of topics with which they have worked. Such a form of representing expertise turned out to be useful in the task of developer recommend. On a dataset of 9700 issues mined from the YouTrack issue-tracker, the accuracy of predictions was improved from 68% to 75%. Usage of interpretable features allowed to keep the models’ predictions interpretable in both developer recommendation and similar projects suggestion tasks.