Dynamic Feature-Aware Attention Fusion of DRN and Machine Learning Models for Software Defect Prediction

Abstract

Software Defect Prediction (SDP) is significant for making sure the software developed is of quality. This is achieved by detecting error-prone components early in the development process. This allows prioritizing testing and quality assurance efforts to minimize maintenance costs, to avoid system failures, and finally to ensure the delivery of a reliable software system. Traditional machine learning (ML) models and ensemble approaches such as soft voting and stacking have shown promise, yet often rely on static weighting strategies that fail to adapt to instance-specific variations. In order to overcome this limitation, they proposed a new Feature-Aware Attention-Based Fusion Model to learn the outputs of multiple base learners and a Deep Residual Network (DRN) based on a dynamic attention mechanism conditioned on the base model predictions and the original input features that allows the model to learn adaptive, instance-specific weights for making the final prediction. The approach incorporates a PSO-CSO hybrid feature selection strategy for dimensionality reduction and uses SMOTE for handling class imbalance. In the fusion layer, a context-aware attention mechanism dynamically integrates the base predictions with the outputs of the Deep Residual Network (DRN). Extensive experiments were conducted on five NASA benchmark datasets (PC5, PC1, KC1, KC2, JM1), and evaluated in comparison to the individual classifiers, soft voting, stacking, and the DRN with static attention. The Empirical evidence suggests that the proposed model outperformed all standard approaches, achieving an accuracy of 0.99, an F1-score of 0.99, a precision of 0.98, a recall of 0.9976, and an impressive ROC (Receiver Operating Characteristic)-AUC (Area Under the Curve) of 0.9992. The results indicate that combining the feature context with the model outputs using attention can form a robust and high-accuracy framework for SDP.