Go to ICLR 2026 Conference homepageSTKAN: Kolmogorov-Arnold Networks for Spatio-Temporal Forecasting
Keywords: Kolmogorov–Arnold Network, Spatial-Temporal Forecasting, Decoupled Spatial-Temporal Modeling, Adaptive Spatial Aggregation
Abstract: Real-world traffic data exhibit intricate, intertwined spatial and temporal dynamics, significantly complicating accurate forecasting. Recent decomposition-based approaches aim to disentangle these complex dynamics into separate spatial and temporal components, facilitating clearer and more effective modeling. However, varying information densities between spatial structures and temporal patterns remain a substantial challenge, potentially leading to inaccurate feature interactions and subsequently degraded forecast performance. Furthermore, existing forecasting models often lack explicit interpretability, obscuring spatial-temporal influences driving predictions. To address these critical challenges, inspired by Kolmogorov-Arnold Networks (KANs), we propose a novel Spatio-Temporal Decomposition Learning architecture (STKAN). The STKAN framework explicitly separates and individually models spatial and temporal dependencies using specialized multi-order KAN modules. It encodes complex input series into spatio-temporal embeddings through an adaptive node-group assignment mechanism. Dedicated spatial and temporal KAN modules independently and robustly capture internode relationships and temporal dynamics at multiple orders, explicitly modeling distinct underlying patterns. Extensive experimental evaluations conducted on widely recognized benchmark datasets convincingly demonstrate that STKAN achieves state-of-the-art forecasting accuracy and interpretability, clearly delineating the critical spatial groupings and essential temporal patterns that significantly influence predictions, thus greatly benefiting real-world intelligent transportation scenarios.
Primary Area: other topics in machine learning (i.e., none of the above)
Submission Number: 5518
Loading