Digital Twin-Driven Adaptive Graph Learning for Explainable Software Reliability Prediction

Abstract

Software reliability prediction is a fundamental activity in software engineering because unreliable software can lead to significant operational, financial, and safety-related consequences. However, existing reliability prediction approaches predominantly rely on static historical datasets and offline analysis, thereby limiting their ability to capture the dynamic behaviour of modern software systems operating under continuously evolving workloads and execution environments. To address these limitations, this paper proposes a novel Digital Twin-Driven Adaptive Graph Learning framework for Explainable Software Reliability Prediction (DTAGL-ESRP). The proposed framework constructs a software digital twin that continuously synchronizes runtime execution traces, operational profiles, software logs, and structural code metrics to create a dynamic virtual representation of the software system. Subsequently, a Temporal Graph Transformer is employed to model inter-module dependencies and capture fault propagation patterns across software components. To maintain prediction robustness under evolving operational conditions, a concept drift detection mechanism is integrated to automatically adapt the learning process whenever significant behavioural changes are observed. Additionally, an explainable artificial intelligence module based on SHAP analysis is incorporated to identify critical reliability-influencing attributes and provide transparent reliability assessment. Experimental evaluation conducted on benchmark NASA dataset demonstrates that the proposed DTAGL-ESRP framework significantly outperforms conventional reliability prediction models in terms of prediction accuracy, adaptability, and interpretability.