software fault prediction

Software fault prediction methods are used to predict fault-prone modules in the early stages of software development. Machine learning techniques are the most common techniques used in software fault prediction. Data dimensionality is one of the problems that affect the performance of machine learning algorithms. Data dimensionality means the existence of irrelevant or redundant features that may mislead the learning algorithm hence decrease its accuracy. Low accuracy of software fault prediction causes late detection of some faulty modules and as a result increases the effort and cost of fixing faults abnormally. Therefore, solving data dimensionality problem is necessary to increase the accuracy of software fault prediction. Researchers use feature selection algorithms for dimensionality reduction. Feature selection algorithms are divided into two types of filter-based feature selection and wrapper-based feature selection algorithms. Wrapper-based algorithms lead to higher accuracy prediction models. In these algorithms we can use different methods to search for the best solutions that best of them is metaheuristic search. Each of the metaheuristic algorithms has some strengths and weaknesses that researchers use combination of these algorithms to address these weaknesses. In this research, to address the weaknesses of each metaheuristic algorithm, a combination of genetic, ant colony and whale optimization algorithm is used as wrapper feature selection. Obviously, the application of early software faults prediction methods before the actual test is one of the effective passive defense techniques in reducing software systems development costs. 19 software projects are used to evaluate proposed method. Comparison of the results with other methods shows that the proposed method outperforms other methods.

Today, due to high costs, it is not possible to perform a comprehensive and complete test on all parts of thesoftware. But if the fault-prone parts are identified before the test, the main focus of the test can be placedon these parts, which leads to cost savings. Identifying fault-prone components is the main purpose ofsoftware fault prediction. A predictive model receives software modules along with their features as inputand predicts which ones are prone to fault. Machine learning techniques are commonly used to constructthese models, the performance of which is highly dependent on the training dataset. Training datasetsusually have many software features, some of which are irrelevant or redundant, and the removal of thesefeatures is done using feature selection methods. In this research, a new method for wrapper-based featureselection is proposed that uses memetic algorithm, random forest technique and a new criterion based onfuzzy inference system. The results show that the proposed fuzzy evaluation criterion has a betterperformance than the existing criteria and improves the performance of feature selection. The final purposeof this research is to achieve a robust model for predicting high performance software faults and thecomparison results show that the proposed model has higher performance than other models.