Predicting the number of defects in a new software version

Area of Science:

• Software Engineering
• Empirical Software Engineering
• Software Quality Assurance

Background:

• Method-level defect prediction is crucial for software quality but remains under-researched.
• Existing research lacks effective strategies for predicting defects in new software versions at the method level.

Purpose of the Study:

• To develop regression models for predicting method-level software defects in new versions.
• To analyze the correlation between defect density, defect velocity, and defect introduction time with method-level defects.
• To investigate the relationship between defect acceleration and defect density.

Main Methods:

• Analysis of historical data from current software versions.
• Construction of regression models using defect density, defect velocity, and defect introduction time.
• Experimentation on the Finding Faults Using Ensemble Learners (ELFF) open-source Java projects (289,132 methods).

Main Results:

• Defect velocity demonstrated a strong correlation (93%) with the number of method-level defects.
• Defect density showed a 60% correlation, while defect introduction time had a -4% correlation.
• A mathematical relationship was identified between defect density and defect acceleration at the method level.

Conclusions:

• Defect velocity is a significant predictor of method-level software defects.
• The proposed approach provides a viable method for predicting defects in new software versions.
• Further research is motivated to explore data preprocessing and classifier performance.