A comparative study of Kolmogorov-Arnold networks and multilayer perceptrons for software defect prediction with SHAP analysis

Software defect prediction is essential for enhancing software quality by detecting software errors early in the development lifecycle. This article presents the usage and a detailed evaluation of Kolmogorov-Arnold Networks (KAN) and Multilayer Perceptrons (MLP) over eleven prevalent software defect prediction datasets, namely CM1, JM1, KC1, KC3, MC1, MC2, MW1, PC1, PC3, PC4, and PC5. The defect prediction models were established and evaluated using five principal performance metrics: Accuracy, Precision, Recall, F1 Score, and Area Under the Curve (AUC). The experimental findings indicate that KAN consistently surpasses MLP in all datasets and metrics. KAN attains an average accuracy enhancement of almost 12% compared to MLP and demonstrates substantial improvements in precision and recall, rendering it more dependable in detecting defected modules. The paired t-test findings indicate that the performance disparities between KAN and MLP are statistically significant (p-value< 0.05) for the majority of metrics. Also SHAP analysis was performed to determine the important rankings of features for all datasets according to the FastKAN model. This study underscores the promise of KAN as a superior alternative to traditional neural networks for software defect prediction. The results highlight the significance of the usage of sophisticated neural network topologies in tackling intricate software engineering challenges.