Scholarlink

<p>Multivariate time series classification is a task that involves extracting features and making categorical distinctions from sequence data composed of multiple time-dependent variables. As the length of time series data increases, the dimensionality grows and the demand for higher classification accuracy rises, traditional multivariate time series classification methods are gradually showing limitations. With the development of deep learning techniques, neural network-based methods have made significant progress in extracting temporal and spatial features, achieving higher classification accuracy. However, existing methods lack a unified framework to capture both multi-scale temporal dependencies and dynamic inter-variable correlations. To address these challenges, this paper proposes an innovative multivariate time series classification method and names it TCGNet. Specifically, TCGNet uses a grouping strategy to divide the multivariate time series data into equally sized groups and applies a multi-head attention mechanism to learn features from each group, uncovering potential temporal patterns across different time scales. Furthermore, TCGNet uses dynamically generated sparse adjacency matrices as inputs in the graph convolutional network to uncover the complex relationships and feature interactions between multiple dimensions of data. Experimental results demonstrate that the proposed classification method achieves superior performance on multiple datasets.</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

Attention-driven hybrid feature learning for multivariate time series classification

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Donghao Lin,
  • Nan Zhang,
  • Hui Huang

摘要

Multivariate time series classification is a task that involves extracting features and making categorical distinctions from sequence data composed of multiple time-dependent variables. As the length of time series data increases, the dimensionality grows and the demand for higher classification accuracy rises, traditional multivariate time series classification methods are gradually showing limitations. With the development of deep learning techniques, neural network-based methods have made significant progress in extracting temporal and spatial features, achieving higher classification accuracy. However, existing methods lack a unified framework to capture both multi-scale temporal dependencies and dynamic inter-variable correlations. To address these challenges, this paper proposes an innovative multivariate time series classification method and names it TCGNet. Specifically, TCGNet uses a grouping strategy to divide the multivariate time series data into equally sized groups and applies a multi-head attention mechanism to learn features from each group, uncovering potential temporal patterns across different time scales. Furthermore, TCGNet uses dynamically generated sparse adjacency matrices as inputs in the graph convolutional network to uncover the complex relationships and feature interactions between multiple dimensions of data. Experimental results demonstrate that the proposed classification method achieves superior performance on multiple datasets.