부산대학교 도서관

In code-based software defect prediction models, the problem of mislabeling often occurs, resulting in reduced model performance and inaccurate decisions. To tackle this issue, this study presents Cleanlab, a confidence learning-based method, for detecting and rectifying mis-labels. Unlike conventional approaches, our emphasis is on code-based defect prediction models. We validate the efficacy of Cleanlab by conducting case studies on 32 versions of nine projects, using three well-known models: CNN, DBN, and LINE-DP. The experimental findings demonstrate substantial enhancements in the robustness and accuracy of our code-based defect prediction model when incorporating Cleanlab.