Scholarlink

<p>In multivariate time series forecasting, effectively modeling both intra-series and inter-series dependencies is key to achieving accurate predictions. Dual clustering along these two dimensions is a commonly used modeling approach; the DUET is a state-of-the-art method among this category. However, it has the following two limitations: (1) its channel correlation metric overemphasizes frequency-domain information while underutilizing temporal information; (2) it struggles to capture multi-scale periodic patterns in complex temporal distributions. To address the two limitations, we propose an improved framework, D-TIME, which consists of two core components: decoupling-guided spatiotemporal-frequency similarity measurement (STSM) and multi-scale pattern decomposition mixture of experts (SMDM). STSM fuses temporal trend and spectral periodicity similarities for cross-domain feature complementarity, while SMDM adaptively integrates multi-scale periodic features via residual decomposition–aggregation. Extensive experiments on eight benchmark datasets, comparing D-TIME with nine representative baseline models, demonstrate that D-TIME achieves superior performance in most scenarios. On average, it reduces the mean squared error (MSE) by approximately 7.8% and the mean absolute error (MAE) by about 5.2%, while maintaining competitive performance across all benchmarks. Our code is available at <a href="https://github.com/wym0924/D-TIME.">https://github.com/wym0924/D-TIME.</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

D-TIME: an improved framework for multivariate time series forecasting

  • 期刊论文

  • 发表日期：

    2026年3月24日

查看全文
  • Yiming Wang,
  • Junhai Zhai

摘要

In multivariate time series forecasting, effectively modeling both intra-series and inter-series dependencies is key to achieving accurate predictions. Dual clustering along these two dimensions is a commonly used modeling approach; the DUET is a state-of-the-art method among this category. However, it has the following two limitations: (1) its channel correlation metric overemphasizes frequency-domain information while underutilizing temporal information; (2) it struggles to capture multi-scale periodic patterns in complex temporal distributions. To address the two limitations, we propose an improved framework, D-TIME, which consists of two core components: decoupling-guided spatiotemporal-frequency similarity measurement (STSM) and multi-scale pattern decomposition mixture of experts (SMDM). STSM fuses temporal trend and spectral periodicity similarities for cross-domain feature complementarity, while SMDM adaptively integrates multi-scale periodic features via residual decomposition–aggregation. Extensive experiments on eight benchmark datasets, comparing D-TIME with nine representative baseline models, demonstrate that D-TIME achieves superior performance in most scenarios. On average, it reduces the mean squared error (MSE) by approximately 7.8% and the mean absolute error (MAE) by about 5.2%, while maintaining competitive performance across all benchmarks. Our code is available at https://github.com/wym0924/D-TIME.