Modeling code quality using machine intelligence
(2022) In Master'sTheses in Mathematical Sciences FMAM05 20221Mathematics (Faculty of Engineering)
- Abstract
- For any company with a software development branch, one of the most important aspects is to write maintainable, understandable, high-quality source code. This will result in fewer work hours to refactor the code if changes are needed. Therefore it’s expensive to keep working with source code of poor quality. The question is, how to measure the quality of source code?
The writers of this thesis approached the problem using machine learn- ing. Firstly two literature studies were conducted, one concerning usable software metrics and one concerning usable machine learning algorithms. Secondly, a large, external, labeled data base was set up and the models were trained on said data. There were different final models set up. Two to focus on... (More) - For any company with a software development branch, one of the most important aspects is to write maintainable, understandable, high-quality source code. This will result in fewer work hours to refactor the code if changes are needed. Therefore it’s expensive to keep working with source code of poor quality. The question is, how to measure the quality of source code?
The writers of this thesis approached the problem using machine learn- ing. Firstly two literature studies were conducted, one concerning usable software metrics and one concerning usable machine learning algorithms. Secondly, a large, external, labeled data base was set up and the models were trained on said data. There were different final models set up. Two to focus on large projects consisting of hundreds of files and two to focus on singular stand-alone files. After examining the models, the ones with the lowest RMSEs on the test set were used to compare the predictions from the models with the opinions of experienced developers.
The final models were built using the algorithms artificial neural network and random forest which both gave similar results. The models for specific files were tested for both front end and back end files where four files in each were ranked according to the final score from the model. The rank was then compared to a ranking of the same files by experienced software developers. The results showed for front end files, that the model agreed with the developers ranking to a large extent. For back end files, there was a wider discrepancy between the opinions of the developers and the predictions made by the model.
The final models were also able to evaluate a large project over time, to evaluate the overall code quality, which also was one thought application of the model. When a project that was rewritten a year ago was continuously evaluated, the code quality improved which was the expected result according to developers in the project. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9092905
- author
- Malmström, Arvid LU and Kindt, Erik LU
- supervisor
- organization
- course
- FMAM05 20221
- year
- 2022
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Machine Learning, Artificial Intelligence, Software Development, Software Metrics
- publication/series
- Master'sTheses in Mathematical Sciences
- report number
- LUTFMA-3486-2022
- ISSN
- 1404-6342
- other publication id
- 2022:E62
- language
- English
- id
- 9092905
- date added to LUP
- 2022-08-05 10:28:02
- date last changed
- 2022-08-05 10:28:02
@misc{9092905,
abstract = {{For any company with a software development branch, one of the most important aspects is to write maintainable, understandable, high-quality source code. This will result in fewer work hours to refactor the code if changes are needed. Therefore it’s expensive to keep working with source code of poor quality. The question is, how to measure the quality of source code?
The writers of this thesis approached the problem using machine learn- ing. Firstly two literature studies were conducted, one concerning usable software metrics and one concerning usable machine learning algorithms. Secondly, a large, external, labeled data base was set up and the models were trained on said data. There were different final models set up. Two to focus on large projects consisting of hundreds of files and two to focus on singular stand-alone files. After examining the models, the ones with the lowest RMSEs on the test set were used to compare the predictions from the models with the opinions of experienced developers.
The final models were built using the algorithms artificial neural network and random forest which both gave similar results. The models for specific files were tested for both front end and back end files where four files in each were ranked according to the final score from the model. The rank was then compared to a ranking of the same files by experienced software developers. The results showed for front end files, that the model agreed with the developers ranking to a large extent. For back end files, there was a wider discrepancy between the opinions of the developers and the predictions made by the model.
The final models were also able to evaluate a large project over time, to evaluate the overall code quality, which also was one thought application of the model. When a project that was rewritten a year ago was continuously evaluated, the code quality improved which was the expected result according to developers in the project.}},
author = {{Malmström, Arvid and Kindt, Erik}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master'sTheses in Mathematical Sciences}},
title = {{Modeling code quality using machine intelligence}},
year = {{2022}},
}