An integrated framework for explainable and uncertainty-aware software defect prediction with adaptive risk scoring and multi-task deep residual learning
Year: 2026
Type: article
Abstract: Software defect prediction (SDP) is a challenging problem in SQA, the majority of the approaches that have been proposed so far focus on this problem as a binary classification problem, without taking into account the risk associated with each defective module. This paper introduces an integrated framework of adaptive risk scoring, Multi-Task deep residual learning with gated feature attention, Monte Carlo Dropout uncertainty quantification, and SHAP-based explainability that is applied simultaneously towards both defect detection and risk classification. This is due to a data-d Single-Task riven parameter λ optimized via a Spearman correlation used to calculate the risk score and SMOTETomek resampling and focal loss to address severe class imbalance. A defect detection ROC-AUC of 0.9581, F1-score of 0.5928, and recall of 0.7616 is achieved by the proposed model on the pc5 dataset; while a macro F1-score of 0.98 and accuracy of 0.99 are achieved by the proposed model on the risk classification task, in which the three risk classes are model-derived prioritisation tiers rather than independently validated defect-severity labels. Monte Carlo Dropout also identifies 25.2% of modules as high uncertainty candidates for manual review, while feature attribution shows that cyclomatic density, comment percentage and global data density are the most important features affecting defect risk.
Source: Scientific Reports
Authors: M. Murali Mohana Kumara Varma, M. Giri
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7221033795