Static code metrics-based deep learning architecture for software fault prediction

Software fault prediction (SFP) refers to the early prediction of fault-prone modules in software development which are susceptible to faults and can incur high development cost. Machine learning (ML)-based classifiers are extensively being used for SFP. Machine learning models utilize handcrafted metrics (or features), i.e., static code metrics for classification of software modules into one of the two categories, i.e., {buggy, clean}. It involves overhead of selecting the most significant features due to the presence of some correlated or non-significant features. With the shifting paradigm of machine learning to deep learning, it is desirable to improve the performance of SFP classifiers to keep pace up with the changing industrial needs. This study proposes a novel model (SCM-DLA-SFP) based on deep learning architecture (DLA) to predict the defects utilizing the static code metrics (SCMs). The defect dataset with SCMs is fed to the input layer of specially designed deep learning model, where the input is automatically conditioned using normalization. Then, the conditioned data pass through dense layers of deep neural network architecture to predict the faulty modules. The study utilizes five datasets from PROMISE repository namely camel, jedit, lucene, synapse and xalan. The proposed model SCM-DLA-SFP exhibits the performance of the average values of 88.01%, 79.83%, and 73.3% for AUC measure, accuracy criteria and F-measure, respectively. The comparison shows that proposed model is better on average than the state-of-the-art DL-based SFP methods by 16.28%, 19.61%, and 18.45% over AUC, accuracy and F-measure, respectively.