Skip to content

DecompKAN: Decomposed Patch-KAN for Long-Term Time Series Forecasting

Home / Papers / DecompKAN: Decomposed Patch-KAN for Long-Term Time Series Forecasting

DecompKAN: Decomposed Patch-KAN for Long-Term Time Series Forecasting

Authors: Naveen Mysore Date: 2026-04-27 Paper ID: openalex:2604.23968

Summary

DecompKAN is a novel time series forecasting architecture that avoids standard attention mechanisms in favor of a pipeline featuring trend-residual decomposition, channel-wise patching, and Kolmogorov-Arnold Network (KAN) layers. By replacing traditional MLP nodes with B-spline-based KAN edges, the model provides inherent transparency, allowing for the direct visualization of latent nonlinear transformations. Evaluation across nine datasets, including the PPG-DaLiA physiological benchmark, shows competitive predictive performance and significant improvements on datasets with smooth temporal dynamics compared to existing transformer-based baselines.

Key Contributions

  • Introduces DecompKAN, an attention-free architecture combining trend-residual decomposition, channel-wise patching, and KAN-based edge functions for time series forecasting.
  • Achieves state-of-the-art or competitive MSE on 20 of 36 controlled-recipe evaluations across 9 benchmarks, outperforming iTransformer on datasets with smooth dynamics like Solar (-17%) and ECL (-10%).
  • Demonstrates that while the architectural pipeline (decomposition, patching) drives forecasting accuracy, the KAN integration uniquely enables visualization of domain-specific latent nonlinearities.

Open Questions & Future Work

Key Concepts

  • kolmogorov-arnold-network: A neural network architecture where learnable functions reside on edges instead of nodes, represented as B-splines to allow for direct visualization of learned nonlinearities.

Archivist Review

I approved the KAN concept and the two open questions because they address the critical intersection of model interpretability and predictive performance in time-series forecasting. The PPG-DaLiA dataset was approved as a high-quality, reusable physiological benchmark, while other architectural components were rejected as they represent standard modular pipelines (patching/decomposition) rather than novel, reusable concepts.

Approved Concepts

  • Kolmogorov-Arnold Network: Central to the paper’s claim of balancing performance with model transparency/inspectability.

Approved Open Questions

  • Characterizing KAN Forecasting Suitability: Understanding the failure modes of KANs is crucial for their adoption in high-stakes scientific forecasting, as it clarifies whether poor performance is a fundamental limitation of B-spline function approximation for non-stationary temporal dynamics or a symptom of current architectural design.
  • Mapping KAN Latents to Features: Bridging the gap between ‘latent-space inspectability’ and ‘feature-space interpretability’ is critical for scientific validity, ensuring that identified patterns correspond to physical phenomena in the raw time series.

Datasets

Metadata & Links

url
https://arxiv.org/abs/2604.23968
paper_id
2604.23968
paper_source
openalex
domain
time-series
tags
architectures
datasets
ppg-dalia
concept_slugs
kolmogorov-arnold-network
dataset_slugs
ppg-dalia
skill
TimeSeriesSkill
created_at
2026-04-30T06:02:37Z