Multilayer Perceptron of Software Complexity Metrics for Explainable Multicollinearity Mitigation and Defect Localization

Abstract

Software defect prediction (SDP) models often neglect the issue of multicollinearity in training sets, particularly among software complexity metrics, which can threaten internal validity and obscure model predictions. This study aimed to investigate the severity of multicollinearity in software complexity metrics, assess their influence on a multilayer perceptron (MLP) model for defect prediction, and evaluate the model's performance with two multicollinearity mitigation techniques. Analysis of the KC1 dataset revealed severe multicollinearity in metrics like lines of code (variance inflation factor [VIF] = 98.08) and cyclomatic complexity (VIF = 399.22). The derived KC1_D dataset of this study exhibited lower VIF values, reducing multicollinearity and improving model stability. The MLP achieved a recall of 0.8452 on the KC1, and a marginal performance improvement on the derived KC1_D, demonstrating resilience against multicollinearity. Principal component analysis (PCA) transformation eliminated multicollinearity by reducing KC1_D attributes to 12 principal components, resulting in perfect classification performance (accuracy, precision, recall, and F1-score = 1.00). Key findings also highlighted feature thresholds influencing software defectiveness, including lines of code + comments (locCodeAndComment) attribute with threshold 0.49) and logical lines of comments (IOComment ≤ −0.33) also contributed significantly to the decision of the MLP model. This study concludes that addressing multicollinearity in software complexity metrics is critical for enhancing the explainability, reliability, and performance of SDP models, particularly through PCA as an effective mitigation technique.