Hybrid Deep Learning Model for Software Defect Classification Using Code2Vec and LinkNet-BiLSTM Score Level Fusion
Srinivasa Rao Katragadda, Sirisha Potluri · International Journal of Electronics and Communication Engineering · 2025
The software package faults stance an important trial, impacting systems’ reliability, functionality, and security. Traditional methods for defect detection, including manual inspections and static tools, are often insufficient for handling large, complex codebases. Current progress in ML and DL provides more robust solutions by identifying complex patterns within data. It presents a hybrid model that combines Code2Vec for feature extraction, Improved LinkNet for spatial data processing, and Bi-LSTM for sequential pattern analysis, leveraging score-level fusion to improve software defect classification. Code2Vec transforms unstructured source code into dense vector representations, capturing critical semantic and syntactic features. Improved LinkNet excels in extracting high-level structural features, while Bi-LSTM captures long-term dependencies in code sequences. The proposed score-level fusion integrates the outputs of these models to harness their complementary strengths, reducing noise sensitivity and enhancing accuracy. The fused scores are passed to a soft-max classifier to foresee a given snippet code that is fault-motionless or non-defect-prone. The final output classifies the software defect into specific categories, if applicable, based on the trained dataset. It demonstrates that the hybrid model outperforms metrics, accuracy, precision, recall, and F1 scores, existing methods in package fault prophecy. The study also highlights the significance of hyperparameter tuning and training large, labelled datasets. This research contributes to scalable, efficient defect detection methods that address real-world challenges in software development, setting a foundation for future improvements in predictive analytics and automated software quality assurance. The proposed model, implemented in Python, enhances classification performance, achieving an accuracy of 92%.