Scholarlink

<p>Traffic forecasting is a typical spatio-temporal data mining problem with broad practical applications. Existing graph-based methods are dedicated to learning dynamic graph structures from traffic data in the time domain to capture spatial correlations in a data-driven manner. However, the time delays and external noise in traffic patterns can weaken the effectiveness of data-driven methods in modeling spatial correlations. Moreover, these methods often focus only on local temporal correlations, failing to effectively capture global temporal correlations. To address these issues, we propose a spatio-temporal Fourier dynamic graph convolutional network (STFDGCN) for traffic forecasting. In the spatial aspect, STFDGCN utilizes the Fourier transform to mitigate the effects of time delays, and introduces sensor identity embedding and time embedding to reduce noise interference. Meanwhile, during the graph convolution process, STFDGCN combines the dynamic graph with adaptive and predefined static graphs to comprehensively capture spatial features. In the temporal aspect, STFDGCN learns local temporal features through dilated causal convolution and introduces a temporal attention module to capture global temporal features. Furthermore, STFDGCN employs a cascading framework to integrate spatial and temporal features. Experiments on four real-world traffic datasets demonstrate that our model is effective and outperforms the baseline models. The source code is available at <a href="https://github.com/hulongfei123/STFDGCN">https://github.com/hulongfei123/STFDGCN</a>.</p>

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

Spatio-temporal Fourier dynamic graph convolution network for traffic forecasting

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • Longfei Hu,
  • Lai Wei,
  • Yeqing Lin

摘要

Traffic forecasting is a typical spatio-temporal data mining problem with broad practical applications. Existing graph-based methods are dedicated to learning dynamic graph structures from traffic data in the time domain to capture spatial correlations in a data-driven manner. However, the time delays and external noise in traffic patterns can weaken the effectiveness of data-driven methods in modeling spatial correlations. Moreover, these methods often focus only on local temporal correlations, failing to effectively capture global temporal correlations. To address these issues, we propose a spatio-temporal Fourier dynamic graph convolutional network (STFDGCN) for traffic forecasting. In the spatial aspect, STFDGCN utilizes the Fourier transform to mitigate the effects of time delays, and introduces sensor identity embedding and time embedding to reduce noise interference. Meanwhile, during the graph convolution process, STFDGCN combines the dynamic graph with adaptive and predefined static graphs to comprehensively capture spatial features. In the temporal aspect, STFDGCN learns local temporal features through dilated causal convolution and introduces a temporal attention module to capture global temporal features. Furthermore, STFDGCN employs a cascading framework to integrate spatial and temporal features. Experiments on four real-world traffic datasets demonstrate that our model is effective and outperforms the baseline models. The source code is available at https://github.com/hulongfei123/STFDGCN.