An Ensemble Learning Method for Software Defect Prediction Targeting Ranking

Software defects pose a significant threat to system stability and user experience. Traditional defect prediction methods often fall short in practical testing scenarios. In response to this, we propose SeDI (Semantics-Enhanced DT-NN Integration), an ensemble learning approach for ordinary software defect prediction. SeDI integrates static software metrics with deep semantic features extracted from abstract syntax trees, leveraging a gradient boosting ranking model and a Multi-layer Perceptron to enhance ranking accuracy. The model's performance is evaluated using FPA and PD20 metrics, demonstrating superior ranking accuracy and improved performance in the top 20% of predicted modules.