Software Defect Prediction via Multi-Channel Convolutional Neural Network

With the growing complexity of modern software, the expense of improving software reliability becomes significant. To reduce costs, Software Defect Prediction (SDP) was proposed to detect hidden bugs a few decades ago. Recent studies demon-strate the superiority of the deep learning-based approach in SDP, such as Convolutional Neural Network (CNN). However, the noise in defect data can still deteriorate their performance. To enhance the performance of the state-of-the-art CNN-based method, we propose an improved SDP framework named Defect Prediction via Multi-Channel Convolutional Neural Network (DP-MC-CNN). The novelty is that DP-MC-CNN extracts sub-trees from the Abstract Syntax Tree, generates multiple paths, and feeds them to a multi-channel CNN, which weakens the noise effect. In evaluation, we assess DP-MC-CNN on seven projects by average F1 score, and it outperforms the state-of-the-art CNN-based method by 3.93%. Furthermore, we validate the effectiveness of the multi-channel approach and reveal the influences of each sub-tree component.