Kolmogorov–Arnold Networks for Accident Risk Prediction at Highway-Rail Grade Crossings

Highway-rail grade crossings, critical intersections where highways and railways meet, pose significant safety challenges, with approximately 2,000 incidents annually in the United States, according to the Federal Railroad Administration (FRA). Existing predictive models, such as those used by the FRA, often rely on rigid assumptions about input transformations, while machine learning models, including multilayer perceptron (MLP), lack interpretability. To address these limitations, this study developed a Kolmogorov–Arnold network (KAN) model to predict accidents at highway-rail grade crossings. The KAN utilized 14 input parameters, including 5 years of accident data, daily train traffic, and safety measures. Multiple basis functions were tested, with Legendre achieving the best overall predictive accuracy and radial basis function (RBF) providing the fastest training times. Across different architectures, compact single-layer models performed well, with the KAN[10] model achieving an R2 of 0.9507 and a test root mean squared error (RMSE) of 0.0075, while the three-layer KAN[10,7,5] model achieved the highest performance (R2=0.9510) at a higher computational cost. This result was compared to other traditional machine learning models, and the KAN outperformed all but two. A rare-event analysis was conducted, and the KAN models outperformed the traditional models. Interpretability analysis revealed warning devices, highway surface type, and accident history as key predictors. Symbolic function surrogates further clarified how these variables influenced the model, with some relationships aligning with engineering expectations. This balance of accuracy and interpretability illustrates KANs as effective tools for enhancing safety decisions at highway-rail crossings, offering valuable information for policy and infrastructure interventions.