University showcase15min
A Two-Stage Machine Learning Framework for File-Level Refactoring Recommendation in Java Projects
The talk presents a two-stage machine learning framework for recommending Java file refactorings: first ranking files likely to be refactored, then predicting refactoring types. It uses Git mining, RefactoringMiner, and software metrics, and highlights that predicted refactorings may not always improve code.
talk.summaryAiDisclaimer
Konstantinos MoukasNetcompany
talkDetail.whenAndWhere
Saturday, November 21, 15:35-15:50
Room 6 - Olympias
Which files should you refactor first, and what changes should you consider? In large Java codebases, identifying refactoring opportunities can demand considerable time and familiarity with the system. This talk explores how machine learning can support that process.
Based on my thesis research at the University of Macedonia, I will present a two-stage framework for file-level refactoring recommendation. The first stage ranks Java files by their predicted likelihood of undergoing refactoring. The second predicts likely refactoring types, helping developers decide where to focus their attention and which changes to investigate.
We will follow the pipeline from mining Git histories and detecting refactorings with RefactoringMiner to extracting software metrics and training predictive models. We will examine how complexity, coupling, cohesion, and their evolution inform recommendations, and how the framework is evaluated on previously unseen projects.
Attendees will learn how software metrics and machine learning can support maintenance decisions, while understanding a key limitation: predicting a refactoring does not establish that it will improve the code. The session is intended for Java developers and anyone interested in software quality and applied machine learning.
Based on my thesis research at the University of Macedonia, I will present a two-stage framework for file-level refactoring recommendation. The first stage ranks Java files by their predicted likelihood of undergoing refactoring. The second predicts likely refactoring types, helping developers decide where to focus their attention and which changes to investigate.
We will follow the pipeline from mining Git histories and detecting refactorings with RefactoringMiner to extracting software metrics and training predictive models. We will examine how complexity, coupling, cohesion, and their evolution inform recommendations, and how the framework is evaluated on previously unseen projects.
Attendees will learn how software metrics and machine learning can support maintenance decisions, while understanding a key limitation: predicting a refactoring does not establish that it will improve the code. The session is intended for Java developers and anyone interested in software quality and applied machine learning.
Konstantinos Moukas
Konstantinos Moukas is a 22-year-old software engineer and a recent graduate of the University of Macedonia’s Department of Applied Informatics, specialising in Computer Science and Technology. He joined Netcompany in July 2025 and currently works as an Associate Engineer, developing backend applications with Java and Spring. His interests include software engineering, refactoring, and the application of machine learning to software development.