During the development of a software system, potential defects and errors can pose irrecoverable risks and damages. Therefore, ensuring the quality and desired performance of software is of utmost importance. One advanced approach to addressing this issue is the use of intelligent and precise methods such as machine learning, which is used for predicting and reducing software defects. The ultimate goal of these methods is to decrease the number of defects in software products and enhance reliability and user satisfaction. Specifically, this research focuses on predicting software defects with an emphasis on reducing functional risks and potential software errors. Given the nature of the problem, which is anomaly detection (identifying defects in software), and upon examining the available dataset in the field of software defect detection, it becomes evident that the distribution of data between defective and non-defective instances is different. The scarcity of data related to defects (anomalies) poses serious challenges for using simple machine learning methods and deep learning. Therefore, in this study, using NASA's available dataset for software defect detection, we first investigate the performance of common machine learning methods and neural networks. Subsequently, by employing deep learning methods and autoencoder networks, we compare their performance with previous methods and aim to improve the detection of software defects. This research utilizes seven different machine learning algorithms for predicting software defects, including Random Forest, Bagging, Multi-layer Perceptron, Radial Basis Function, Naive Bayes (Gaussian), Naive Bayes (Multinomial), Support Vector Machine, Artificial Neural Networks, and Autoencoder Networks. These algorithms have been implemented on NASA's available datasets to predict software defects. In this research, based on the findings and results of eva‎luating the performance of networks and considering the existence of two imbalanced data classes, it was evident that classical machine learning methods face challenges in detecting defective data. However, in contrast, autoencoder networks can easily overcome challenges in anomaly detection within an imbalanced dataset. Therefore, this study has constructed a classifier for detecting rare events using autoencoder networks. This method serves as a novel solution that significantly improves accuracy in detecting rare events. In summary, the proposed autoencoder model achieved better performance compared to all other implemented methods in this research, reaching an accuracy of 97% on the MC1 dataset and 96% on the KC1 dataset. Additionally, through confusion matrix analysis, it was observed that the autoencoder model detected defective data with much higher accuracy and fewer false positives and negatives, demonstrating its superior performance compared to other methods.

الهام محمودزاده

محمدرضا احمدزاده , شيرين بقولي زاده