STKAN uses Kolmogorov-Arnold Networks to beat traffic forecasting benchmarks

Real-world traffic data is notoriously difficult to forecast due to heterogeneous spatial correlations and nonlinear temporal dynamics. While prior work has focused on increasingly complex graph, attention, and decomposition architectures, the role of the underlying nonlinear function approximator has been largely overlooked. The new paper introduces STKAN, a spatio-temporal forecasting architecture that brings Kolmogorov-Arnold Networks into the mix. KANs, which learn nonlinear functions by composing univariate functions, are applied here as Taylor-polynomial expansions within both spatial and temporal token mixing layers. The architecture first builds high-level spatial representations using a learnable soft node-group assignment, performs group-wise spatial mixing, and then models temporal dependencies over the compressed sequence. Spatial and temporal self-attention layers are added to capture long-range interactions.

Experiments on five traffic forecasting benchmarks demonstrate that STKAN achieves competitive performance and consistently outperforms an equivalent MLP-based variant in the tested settings. These results suggest that the choice of nonlinear function approximator is not merely a detail—it can serve as a useful complement to architectural improvements in spatio-temporal forecasting. The work opens the door for further exploration of KAN-based designs in domains like weather, energy, and mobility prediction. For practitioners, it means that swapping out standard MLPs for KAN modules could yield accuracy gains without overhauling the entire model architecture.