Heterogeneous cross-project defect prediction with multiple source projects based on transfer learning.

Area of Science:

• Software Engineering
• Machine Learning
• Data Mining

Background:

• Cross-project defect prediction (CPDP) traditionally assumes identical metrics between source and target projects.
• Existing heterogeneous CPDP methods are limited to single or homogeneous multiple source projects, restricting data availability.
• This limitation hinders the practical application of defect prediction models.

Purpose of the Study:

• To propose Heterogeneous Defect Prediction with Multiple source projects (HDPM) for improved defect prediction.
• To leverage transfer learning for knowledge transfer across heterogeneous projects.
• To enhance the scope of data acquisition for defect prediction models.

Main Methods:

• HDPM utilizes transfer learning to bridge the domain gap between heterogeneous source and target projects.
• A projective matrix is constructed to align the data distributions of source and target projects.
• The method enables the use of multiple, dissimilar source projects for training defect prediction models.

Main Results:

• Experiments on 14 projects across four datasets demonstrate HDPM's superior performance over existing CPDP methods.
• HDPM achieves performance comparable to or better than within-project defect prediction.
• The approach effectively extends the range of usable data for defect prediction.

Conclusions:

• HDPM successfully addresses the limitations of prior heterogeneous cross-project defect prediction methods.
• The use of multiple heterogeneous source projects significantly enhances defect prediction capabilities.
• This advancement promotes wider applicability of software defect prediction in software engineering.