Software defect prediction based on a multiclassifier with hyperparameters: Future work

Background: Software defects represent a critical challenge for the technology industry, as delayed detection can significantly compromise the quality of the final product. Moreover, these defects lead to substantial increases in costs associated with error correction and additional testing, cause delays in established timelines, impact the reputation of organizations, and heighten security-related risks. Objective: Because data balancing is a problem in software defect prediction, a method is proposed and 4 stacking models based on hyperparameters for improve prediction and performance. Furthermore, a web interface was developed with the best model proposed in this paper. Methods: Firstly, the dataset utilized was source from Kaggle (Software Defect Prediction), which consisted of 10,886 software defect records and 22 attributes. Therefore, the article consists up of the following phases: Cleaning and Preprocessing; Describe the data; Training and testing data; Cross-validation; Model calibration; and modelling and evaluation. Additionally, the different models proposed for predicting software defects were compared using hyperparameter-based stacking, considering the performance evaluation metrics. Results: Stacking 3A (Gradient Boosting) using Oversampling in the testing achieved a higher Accuracy (95.64 %), Sensitivity (95.65 %), F1-Score (95.65 %) and Precision (95.64 %), while the same model together with Random Forest (without balancing) and Bayesian networks using Oversampling achieved the best ROC Curve (98.00 %). Conclusions: By implementing 4 hyperparameter-based Stacking models, it helps to perform early prediction of software defects and greater accuracy, while decreasing the number of potential future problems. Therefore, the combined method demonstrated enhanced accuracy in predicting software defects, surpassing the performance of the individual algorithms employed.