Instance filtering on Cross-project defect prediction with meta-learning

• Grant number: 16/09315-2
• Support Opportunities: Scholarships abroad - Research Internship - Doctorate
• Start date: June 23, 2016
• End date: December 22, 2016
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Adenilso da Silva Simão
• Grantee: Faimison Rodrigues Porto
• Supervisor: Maria Emilia Xavier Mendes
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Institution abroad: University of Oulu, Finland
• Associated to the scholarship: 13/01084-3 - Investigating Software Testing from the Perspective of Complex Networks Theory, BP.DR

Abstract

Defect prediction models can be a good tool to organize the test resources of a project. However, not all companies maintain an appropriate dataset of defects, forcing them to build it from known external projects. This approach, called Cross-project Defect Prediction (CPDP), solves the lack of defect data, although introduces heterogeneity on data. This heterogeneity can compromises the performance of CPDP models. Recently, filtering methods were proposed in order to decrease the heterogeneity of data by selecting the most similar instances from the training dataset. The similarity between instances is calculated based on the project features. We have investigated if the use of features subsets as similarity measures (IFFS) can improve the performance of filtering methods. The results do not indicate an IFFS method with general better performance. Instead, the most efficient IFFS method for a project can vary according to its properties. We propose to investigate the use of meta-learning to predict the most efficient IFFS method for a project. We also propose to investigate the use of global network metrics as meta-features. A meta-dataset composed by relevant meta-features can provide the necessary knowledge to predict the most efficient IFFS method for a project and, consequently, improve the defect prediction performance of CPDP models. (AU)