Software Fault Prediction With an Iterative Fuzzy Logic System Considering Interpretability With Imbalanced Datasets

Users expect software to be error-free; however, preventing faults in software while being developed is difficult. Although predicting faults in software is arduous, it radically helps to improve the software quality. Due to the complexity of software, time, and budget limitations, such prediction helps to deliver more robust and error-free software with lower expenses. This paper introduces an iterative method based on fuzzy systems and machine learning to predict software faults. High interpretability, transparency, balancing data, and finding the best interval for converting numerical features to fuzzy features are basic challenges for predicting software faults. The proposed framework is split into four phases. In the first phase, the crisp inputs are converted to fuzzy sets. In the second phase, a membership function is constructed using triangular fuzzy sets. In the third phase, training data are balanced, and fuzzy rules are generated. In the last phase, the similarity of inputs with the rules’ antecedents is calculated, and the fired rules are aggregated to label the test data. Eclipse, Promise, and Travis repositories are evaluated with the proposed method. The calculated AUC of the proposed method on Promise, Travis, and Eclipse datasets are, respectively, equal to 89%, 62% and 87%, which are comparable to the results obtained by deep learning methods but with higher interpretability and transparency.