Evaluation of Machine Learning Algorithms for Outlier Detection in Clustered Code Fragments

Many software systems are designed to be long-lived due to the costs involved in developing new systems. Changes in these systems are inevitable due to constant modifications in requirements that are necessitated by the constantly changing nature of the business environment or detection of faults.To adapt their software to all these changing requirements developers use several tools, i.e. recommendations systems. To create recommendation patterns, the recommendation system searches for groups of similar code changes in software archives using syntactical similarities. As this search is a heuristic approach the groups contain outliers that prevent the generation of many useful patterns. In this thesis, we device algorithms that work hand in hand with various classification algorithms to identify outliers in the initial groups generated by SIFE. The goal is to improve the final recommendations presented to the developer. For the improved classification we first use manifold learning algorithms to map our data to a three dimensional space. We then use the 3D coordinates of generalizable groups as our feature vectors and train group-specific, project-specific and global classifiers. With these classifiers we make changes to the initial groups. We evaluate the results of the changed groups with the Disruptor, Retrofit, Picasso, Flym, Android Chart and Android Universal Image Loader software repositories. Our approach results in an improvement of up to 36% in the repositories. Joint degree with Friedrich-Alexander-Universität Erlangen-Nürnberg.

Subject

Machine Learning
Recommender Systems

To reference this document use:

http://resolver.tudelft.nl/uuid:2cc644d3-c95c-4331-8b07-445d491f697f

Part of collection

Student theses

Document type

master thesis

Rights

Files

PDF

COSSE_THESIS_JWafula.pdf

1.48 MB

Close viewer