Software defect prediction using hybrid model (CBIL) of convolutional neural network (CNN) and bidirectional long...

Area of Science:

• Software Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Software quality improvement is a key focus in the software industry.
• Proactive defect prediction aids developers and testers in early defect identification, saving time and resources.
• Traditional defect prediction models often overlook source code semantics, relying on superficial features.

Purpose of the Study:

• To propose a novel hybrid model, CBIL, for predicting defective areas in source code.
• To leverage Abstract Syntax Tree (AST) token semantics for improved defect prediction accuracy.
• To enhance software quality through more effective and proactive defect identification.

Main Methods:

• CBIL extracts Abstract Syntax Tree (AST) tokens and converts them into dense vectors using mapping and word embedding.
• A Convolutional Neural Network (CNN) is employed to extract semantic features from AST tokens.
• Bidirectional Long Short-Term Memory (Bi-LSTM) is utilized to retain crucial features and improve prediction accuracy.

Main Results:

• The CBIL model demonstrated a 25% improvement in average F-measure compared to CNN.
• CBIL achieved an 18% improvement in average AUC compared to Recurrent Neural Network (RNN).
• The model was evaluated on seven open-source Java projects from the PROMISE dataset.

Conclusions:

• The proposed CBIL model significantly enhances software defect prediction accuracy by incorporating code semantics.
• Hybrid models combining CNN and Bi-LSTM offer superior performance over traditional and単一 deep learning approaches.
• CBIL provides a promising approach for proactive defect prediction, contributing to improved software quality.