Predicting Requirement Change Using Bayesian Networks on Historical Traceability Data

• Titilope A. BANJOKO

  Author

• Bilkisu L. MUHAMMAD-BELLO

  Author

Requirement change is still a challenge in the evolution of software systems. Machine learning techniques have been widely used in defect prediction and traceability recovery. However, there is relatively less research on probabilistic modelling of short-term requirement change continuation based on graphical structures. This paper explores whether traceability-derived metrics can be used to predict requirement change using window-level instance. The goals of the research were to: (i) build a temporal window-based dataset from trace data in the repository; (ii) derive structural and historical evolution metrics; (iii) perform Bayesian Network structure and parameter learning; and (iv) assess predictive accuracy and explainability. Based on the Eclipse Java Development Toolkit (Eclipse JDT) repository, 4,500 commits were retrieved and associated with bug-level trace data. A 14-day sliding window approach has been employed to aggregate metrics like churn rate, number of authors, and files modified. Log-transformed and discretised features were modelled using Bayesian Networks with Hill-Climbing search for structure learning while parameter estimation was carried out with BDeu prior with Bayesian estimation. The mean ROC-AUC value was 0.610 with a standard deviation of 0.014 over five stratified folds. Statistical testing showed that this was a significant improvement over random guessing (p < 0.001). The inferred network topology showed interpretable probabilistic dependencies between the intensity of structural modification and the continuation of change. The results show that there is a measurable probabilistic pattern in the continuation of window-level instance changes in traceability data, although the predictive power is still moderate. The paper concludes that probabilistic graphical modelling is a useful approach for requirement evolution analysis and suggests future work in semantic feature modelling, multi-project validation, and dynamic probabilistic modelling.

• FJET_21_94_94

Copyright (c) 2026 FUDMA Journal of Engineering and Technology

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.