STKAN: Kolmogorov-Arnold Networks for Spatio-Temporal Forecasting

Short summary

STKAN introduces Taylor-polynomial Kolmogorov-Arnold Network modules into spatio-temporal forecasting for traffic data, replacing standard MLP-based function approximators. It uses learnable soft node-group assignment for spatial mixing followed by temporal modeling over compressed sequences, with self-attention for long-range interactions. Experiments on five traffic benchmarks show competitive performance, suggesting nonlinear function approximator design complements architectural improvements in spatio-temporal forecasting.

• •Introduces KAN modules into spatio-temporal forecasting for improved nonlinear function approximation
• •Uses learnable soft node-group assignment for spatial representation and temporal modeling
• •Achieves competitive results on five traffic forecasting benchmarks

Generated with AI, which can make mistakes.