Kolmogorov-Arnold Networks Enhance Spatio-Temporal Forecasting

Spatio-temporal forecasting, particularly for complex data like traffic, faces challenges due to heterogeneous spatial correlations and nonlinear temporal dynamics. While existing methods have focused on sophisticated architectures like graphs and attention mechanisms, the role of the underlying nonlinear function approximator has been less explored. This research proposes STKAN, a new architecture that incorporates Taylor-polynomial Kolmogorov-Arnold Network (KAN) modules. These KAN modules are integrated into both spatial and temporal token mixing processes. STKAN first creates high-level spatial representations using a learnable soft node-group assignment, then applies group-wise spatial mixing, and finally models temporal dependencies. The model also utilizes spatial and temporal self-attention layers to capture long-range interactions. Experiments on five traffic forecasting benchmarks demonstrate that STKAN achieves competitive performance, often outperforming MLP-based variants. This suggests that enhancing nonlinear function approximators can significantly complement architectural design in spatio-temporal forecasting.