GraphKAN: An Efficient Graph Kolmogorov Arnold Networks for Traffic Forecasting
Wenzhu Zhao, Guan Yuan, Rui Bing, Xiao Liu, Guixian Zhang · 2024
Traffic forecasting is an essential task for the progression of smart cities. Graph Neural Networks (GNNs) based traffic forecasting models can effectively model the complex spatio-temporal dependencies in traffic data and have become the most widely used traffic forecasting method in recent years. However, these methods suffer from high computational complexity, cumbersome structures, and optimization difficulties. To address the above issues, we propose an Efficient Graph Kolmogorov Arnold Networks for Traffic Forecasting (GraphKAN). Specifically, we first design patch processing to delineate the original traffic sequences to reduce the computational overhead of the model. Then we propose an adaptively generated dynamic graph network to capture the dynamic spatial dependence in the traffic data efficiently. We utilize Kolmogorov Arnold Networks with adaptive activation functions to capture complex relationships and nonlinear patterns in time series, which replace traditional linear weights with spline-parametrized univariate functions, significantly reducing the number of parameters and improving interpretability. Finally, we validate the effectiveness and efficiency of GraphKAN via various experiments. The results show that our GraphKAN is advanced in terms of computational speed, number of model parameters, and prediction performance.