A comparative study of software defect binomial classification prediction models based on machine learning

As information technology continues to advance, software applications are becoming increasingly critical. However, the growing size and complexity of software development can lead to serious flaws resulting in significant financial losses. To address this issue, Software Defect Prediction (SDP) technology is being developed to detect and resolve defects early in the software development process, ensuring high software quality. As a result, SDP research has become a major focus for academics worldwide. This study aims to compare various machine learning-based SDP algorithm models and determine if traditional machine learning algorithms affect SDP outcomes. Unlike previous studies that aimed to identify the best prediction model for all datasets, this paper constructs SDP superiority models separately for different datasets. Using the publicly available ESEM2016 dataset, 13 machine learning classification algorithms are employed to predict software defects. Evaluation indicators such as Accuracy, AUC(Area Under the Curve), F-measure, and Running Time(RT) are utilized to assess the performance of the classification algorithms. Due to the serious class imbalance problem in this dataset, 10 sampling methods are combined with the 13 machine learning algorithms to explore the effect of sampling techniques on the performance of traditional machine learning classification models. Finally, a comprehensive evaluation is conducted to identify the best combination of sampling techniques and classification models to construct the final dominant model for SDP.