Scholarlink

Time series forecasting faces challenges from non-stationarity, complex dependencies, and computational demands. Temporal Kolmogorov-Arnold Networks (TKANs) have emerged as a promising alternative to traditional deep learning models, leveraging learnable univariate functions for improved parameter efficiency and interpretability. However, a systematic comparison of TKAN variants is lacking. This paper presents the first comprehensive evaluation of state-of-the-art TKAN models, including TimeKAN, RMoK, and MT-KAN, across large-scale benchmark datasets. We identify a shared limitation in modeling inter-variable dependencies for multivariate forecasting and introduce HiPPO-KAN-GNN, a novel extension incorporating Graph Neural Networks to address this issue. Experiments on a high-performance computing cluster demonstrate consistent accuracy improvements with our model. Additionally, we conduct a sustainability analysis, measuring energy consumption and carbon footprint, offering critical insights for green AI deployment. This study provides a holistic view of TKANs’ performance, scalability, and environmental impact, advancing both their practical application and sustainable machine learning research.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

HiPPO-KAN-GNN: A Novel Graph Neural Networks-Based Extension to Temporal Kolmogorov-Arnold Networks for Multivariate Time Series Forecasting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Devvrat Joshi,
  • Pancham Shukla

摘要

Time series forecasting faces challenges from non-stationarity, complex dependencies, and computational demands. Temporal Kolmogorov-Arnold Networks (TKANs) have emerged as a promising alternative to traditional deep learning models, leveraging learnable univariate functions for improved parameter efficiency and interpretability. However, a systematic comparison of TKAN variants is lacking. This paper presents the first comprehensive evaluation of state-of-the-art TKAN models, including TimeKAN, RMoK, and MT-KAN, across large-scale benchmark datasets. We identify a shared limitation in modeling inter-variable dependencies for multivariate forecasting and introduce HiPPO-KAN-GNN, a novel extension incorporating Graph Neural Networks to address this issue. Experiments on a high-performance computing cluster demonstrate consistent accuracy improvements with our model. Additionally, we conduct a sustainability analysis, measuring energy consumption and carbon footprint, offering critical insights for green AI deployment. This study provides a holistic view of TKANs’ performance, scalability, and environmental impact, advancing both their practical application and sustainable machine learning research.