Scholarlink

<p>Multivariate Time Series Analysis aims to utilize past time series data to predict future trends. In recent years, Multi-Layer Perceptron methods have gained significant attention due to their lightweight structure and superior performance. Some models based on Transformer architectures and Convolutional Neural Networks also use linear layers for decoding. However, due to the differences among various channels in time series, a single linear model restricts the model’s capability to effectively predict different channels. As a representative linear model, DLinear allows the allocation of parameter-independent linear layers for different channels to enhance model performance. However, this approach may lead to issues such as insufficient training of the linear layers and a significant increase in model complexity when the number of channels is large. Therefore, we propose a Grouped Parameter-Independent Linear Model named GLinear. By calculating the correlations and applying community detection algorithms to group the correlated variables, we apply different linear layer structures to the grouped channels. GLinear allows for the explicit control of correlations within groups and requires computation only once for different datasets, thus achieves an effective balance in effectiveness and complexity. Experiments on seven real-world datasets demonstrate that GLinear significantly outperforms classical linear models while maintaining lower complexity.</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

GLinear: multivariate time series forecasting method towards controllable and efficient grouping

  • 期刊论文

  • 发表日期：

    2025年8月8日

查看全文
  • Gaoxiang Zhao,
  • Xiaoqiang Wang

摘要

Multivariate Time Series Analysis aims to utilize past time series data to predict future trends. In recent years, Multi-Layer Perceptron methods have gained significant attention due to their lightweight structure and superior performance. Some models based on Transformer architectures and Convolutional Neural Networks also use linear layers for decoding. However, due to the differences among various channels in time series, a single linear model restricts the model’s capability to effectively predict different channels. As a representative linear model, DLinear allows the allocation of parameter-independent linear layers for different channels to enhance model performance. However, this approach may lead to issues such as insufficient training of the linear layers and a significant increase in model complexity when the number of channels is large. Therefore, we propose a Grouped Parameter-Independent Linear Model named GLinear. By calculating the correlations and applying community detection algorithms to group the correlated variables, we apply different linear layer structures to the grouped channels. GLinear allows for the explicit control of correlations within groups and requires computation only once for different datasets, thus achieves an effective balance in effectiveness and complexity. Experiments on seven real-world datasets demonstrate that GLinear significantly outperforms classical linear models while maintaining lower complexity.