TRGNet: a deep transfer learning approach for software defect prediction

Abstract

Software defect prediction (SDP) aims to automatically locate defective modules to find bugs and prioritise testing efforts. Researchers are now shifting into semantic features in order to develop predictive models for accurate prediction by using deep learning. But the source code conversion into the semantic feature fails to capture the essential features and correlation. This often degrades the performance of the prediction model. However, well-known authors have already shown the importance of software module metrics for software defect prediction. To take the advantage of software metrics via deep transfer learning in this paper, software module metrics are transformed into images. We proposed the TRGNet model, which extracts transferable features from source projects using pre-trained GoogLeNet and consolidates with a meta-estimator to minimize the divergence in sample distributions between projects. In this model, we feed the transformed image file of software modules to train it for within-project defect prediction (WPDP) and cross-project defect prediction (CPDP). The experimental results with AlexNet, ResNet, SqueezeNet, and other state-of-the-art models indicate that the proposed TRGNet model significantly improves the state-of-the-art defect prediction task by 13.31 % in WPDP and 16.88 % in CPDP scenarios. Moreover, the computational cost analysis reveals that TRGNet significantly reduces memory utilization while maintaining competitive training and inference times compared to other deep learning models, making it a highly efficient and scalable approach for SDP.