Research of Software Defect Prediction Model Based on Complex Network and Graph Neural Network.

Area of Science:

• Computer Science
• Software Engineering
• Artificial Intelligence

Background:

• Traditional software defect prediction models primarily analyze code features, neglecting inter-module relationships.
• Understanding software module dependencies is crucial for accurate defect prediction.

Purpose of the Study:

• To propose a novel software defect prediction framework using graph neural networks (GNNs) from a complex network perspective.
• To leverage the relationships between software modules for more effective defect prediction.

Main Methods:

• Representing software as a graph with classes as nodes and dependencies as edges.
• Employing community detection algorithms to partition the software graph into subgraphs.
• Utilizing an improved GNN model to learn node representation vectors for defect classification.

Main Results:

• The proposed GNN framework demonstrated significant improvements in accuracy, F-measure, and Matthews correlation coefficient (MCC) on the PROMISE dataset.
• Both spectral and spatial domain graph convolution methods within the GNN showed performance gains over benchmark models.
• Average metric improvements ranged from 6.3% to 17.5% depending on the specific metric and convolution method used.

Conclusions:

• The graph neural network approach effectively captures inter-module dependencies, leading to enhanced software defect prediction.
• This complex network perspective offers a promising direction for advancing software defect prediction methodologies.