Semantic and traditional feature fusion for software defect prediction using hybrid deep learning model

Area of Science:

• Software Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Traditional software defect prediction models rely on code metrics, which often fail to capture essential semantic nuances.
• Semantic features, while insightful, lack statistical metrics like code size and complexity, limiting their standalone effectiveness.
• Integrating both feature types is crucial for robust defect prediction.

Purpose of the Study:

• To propose a novel defect prediction model that synergistically combines traditional and semantic software features.
• To address the limitations of using single-feature types in software defect prediction.
• To enhance the accuracy and reliability of predicting software defects.

Main Methods:

• A hybrid deep learning model, Convolutional Neural Network (CNN)-Multilayer Perceptron (MLP), was developed.
• CNN processed semantic features extracted from Abstract Syntax Trees (ASTs) using Word2vec.
• MLP processed traditional features, with outputs integrated for final defect prediction.

Main Results:

• The proposed CNN-MLP model demonstrated significant enhancements in software defect prediction performance.
• Extensive experiments on open-source projects validated the model's effectiveness.
• The CNN-MLP approach outperformed existing methods in both non-effort-aware and effort-aware scenarios.

Conclusions:

• Integrating traditional and semantic features via a hybrid deep learning approach is highly effective for software defect prediction.
• The CNN-MLP model offers a superior solution for identifying software defects compared to existing techniques.
• This research provides a more reliable method for software engineers to allocate resources for high-quality software releases.