LRNN: A Formal Logic Rules-Based Neural Network for Software Defect Prediction

Class imbalance in deep learning systems can lead to biased models due to the unbalanced distribution of data classes in the training dataset. When such biased models are adopted for software defect prediction (SDP), the prediction precision will be significantly affected. To improve the mechanism of existing related deep learning models for SDP, we propose a new neural network, called Logic Rules Neural Network (LRNN), to address the class imbalance problem. LRNN is characterized by utilizing formal logic rules on the association between code metrics to improve the efficiency of deep learning. We describe how relevant formal logic can be identified and selected and how it can be utilized in the learning process. We evaluate the performance of our approach by conducting some experiments using the NASA, PROMISE, AEEEM repositories. The results demonstrate a significant improvement over several existing algorithms.