Abstract Graph Neural Networks (GNNs) have emerged as effective models for analyzing complex and structured data in software engineering tasks, particularly issue categorization and bug triage. This study investigates the performance of Prior Structural Information GNN (PSI-GNN) on the open-source TAWOS dataset containing over 500 000 Agile project issues. Our pipeline includes an ETL (Extract–Transform–Load) process and semantic descriptor extraction using Term Frequency–Inverse Document Frequency (TF–IDF) and Porter Stemmer to transform issue descriptions into rich feature representations. PSI-GNN is optimized using the Taguchi orthogonal-array method, achieving exceptional eco-efficiency with an F1 score of 94.86% and a runtime of 611 seconds, compared to 40567 seconds using Optuna. Model interpretability is enabled through GNNExplainer, which identifies high-impact keywords for the four target issue types (, , , ), while SHAP analysis highlights feature importance. t-SNE visualizations further illustrate the model’s ability to form distinct clusters of issue types, reflecting PSI-GNN’s capability to capture dependencies within Agile workflows. Overall, the results demonstrate that PSI-GNN provides a computationally efficient and interpretable solution for real-time issue tracking and prioritization, supporting seamless integration with graph databases, feature engineering pipelines, and Agile project management tools.
Ranković et al. (2026) studied this question.