PSUT Portal Project Title: Multi-Modal Mamba-Transformer and Knowledge Graph Framework for Just-in-Time Cross-Project Software Defect Prediction
Responsible Researcher: Prof. Mohammad Azzeh
Project Description:
Modern software development under CI/CD pipelines demands rapid, reliable defect detection at the point of code submission. While Just-in-Time Software Defect Prediction (JIT-SDP) enables commit-level risk identification, existing cross-project approaches remain limited by heavy reliance on manual feature engineering, shallow semantic understanding of code and commit text, poor generalization across heterogeneous projects, and an inability to unify diverse information sources (code, text, developer activity, dependencies) into a single coherent model. This project proposes a novel hybrid framework for Just-in-Time Cross-Project Software Defect Prediction (JIT-CPDP) that integrates NLP for textual artifacts, code transformer models for semantic code representation, Mamba neural networks for efficient long-sequence modeling, and graph neural networks combined with knowledge graph representation for capturing relational software structure. By fusing these multi-modal representations, the framework aims to advance prediction accuracy, interpretability, and transferability across diverse open-source repositories, ultimately supporting more proactive and scalable software quality assurance in fast-paced development environments.
Research Questions:
In this project, the following research questions will be investigated (these are too wide, and need to be defined/selected later on):
- How can multi-modal representations—combining source code semantics, commit-message text, and structural/relational software knowledge—be effectively fused to improve commit-level defect prediction accuracy compared to single-modality approaches?
- To what extent can Mamba-based architectures address the computational and long-range dependency limitations of Transformer models in modelling sequential software change history, without sacrificing predictive performance?
- How does the integration of graph neural networks and knowledge graph representations (capturing relationships among commits, developers, files, and dependencies) improve the transferability of defect prediction models across projects with differing structures and development practices?
- What is the comparative effectiveness of the proposed hybrid framework against state-of-the-art baselines (traditional ML, LSTM/GRU, Transformer-based, and GNN-based models) in terms of accuracy, generalization, and computational efficiency across benchmark datasets (e.g., Apache, Eclipse, Mozilla)?
Application Requirements:
The applicant must hold a BSc in IT related major (Data Science, Computer Science, Software Engineering, etc.), have a very good level of mathematics and strong Python programming skills, and meet the Emerging Researchers Scholarship application requirements.
Application Procedure:
Interested applicants should submit their CV along with all required documents to the following email address:
m.azzeh@psut.edu.jo
For more information about the program, please visit: Emerging Researcher Scholarship
Important Dates:
Deadline for scholarship applications: 5 September 2026
Start of the First Semester 2026/2027: 4 October 2026