Comparative Analysis of KAN and MLP Models for Feature Selection Based on Network Weights and SHAP Explanations

Abstract

Breast cancer is one of the most prevalent cancers among women, and its early detection plays a crucial role in improving treatment outcomes. Machine learning has been widely adopted to support breast cancer prediction by developing predictive models from medical data. However, not all features in a dataset contribute equally to the prediction task; irrelevant or redundant features may degrade model performance and increase computational complexity. Therefore, feature selection is a critical preprocessing step for improving both predictive accuracy and computational efficiency. In this study, we investigate feature selection for breast cancer prediction using two neural network models: the Multi-Layer Perceptron (MLP) and the Kolmogorov–Arnold Network (KAN). First, we evaluate the feature importance scores generated independently by each model. We then propose a simple aggregation strategy that combines the importance scores from both models by averaging them, resulting in a more robust and reliable feature ranking. Feature importance is computed using two complementary approaches: the weights connecting the input layer to the first hidden layer and SHAP (SHapley Additive exPlanations) values, which quantify the contribution of each feature to the model’s predictions. The selected features are evaluated on the Wisconsin Breast Cancer dataset using four classification algorithms: K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Decision Tree, and Random Forest. The experimental results demonstrate that the proposed aggregation strategy, which combines feature importance scores from MLP and KAN, produces more consistent feature rankings and improves classification performance compared with using either model individually.