Scholarlink

This paper presents CycleKAN, a novel framework integrating Kolmogorov-Arnold Networks with explicit cycle modeling to address urban traffic forecasting challenges. With increasing urban mobility demands, accurate and efficient traffic prediction has become critical for intelligent transportation systems. CycleKAN introduces three key innovations: a recurrent cycle module with learnable phase modulation and adaptive amplitude control, Taylor-KAN with adaptive order selection for stable polynomial expansion, and dual-path instance normalization bridging time and frequency domains. These components effectively tackle fixed periodicity assumptions, context-blind decomposition, and numerical instability in long-term forecasting. Comprehensive evaluations on six benchmark datasets demonstrate significant performance improvements, achieving lowest MAE while delivering 86.7% latency reduction through parameter sharing and adaptive approximation techniques. The framework performs effectively in resource-constrained edge computing environments with only 3.2 s latency making it highly suitable for real-world traffic management applications.

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

CycleKAN: Integrating Kolmogorov-Arnold Networks with Explicit Cycle Modeling for Efficient Urban Traffic Forecasting

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yuntian Hou,
  • Di Zhang,
  • Qiang Niu

摘要

This paper presents CycleKAN, a novel framework integrating Kolmogorov-Arnold Networks with explicit cycle modeling to address urban traffic forecasting challenges. With increasing urban mobility demands, accurate and efficient traffic prediction has become critical for intelligent transportation systems. CycleKAN introduces three key innovations: a recurrent cycle module with learnable phase modulation and adaptive amplitude control, Taylor-KAN with adaptive order selection for stable polynomial expansion, and dual-path instance normalization bridging time and frequency domains. These components effectively tackle fixed periodicity assumptions, context-blind decomposition, and numerical instability in long-term forecasting. Comprehensive evaluations on six benchmark datasets demonstrate significant performance improvements, achieving lowest MAE while delivering 86.7% latency reduction through parameter sharing and adaptive approximation techniques. The framework performs effectively in resource-constrained edge computing environments with only 3.2 s latency making it highly suitable for real-world traffic management applications.