基于图卷积神经网络的时间序列分类研究

Research on Time Series Classification Based on Graph Convolutional Networks

摘要: 本文提出了一种基于图卷积神经网络(GCN)的时间序列分类方法，通过融合余弦相似度与动态时间规整(DTW)的混合度量构建时序关系图，并在特征提取上增加网络特征提取模块，有效解决了传统方法在特征表达和相似度度量方面的局限性。在模拟数据集和股票价格数据等真实场景中，本方法在分类指标上有良好的效果，多数分类准确率稳定在83%以上，部分类别接近90%。这验证了图神经网络在捕捉时间序列时空依赖性方面的优势，为时间序列分析提供了新的技术路径。

Abstract: This paper presents a time series classification method based on Graph Convolutional Networks (GCN). The method constructs a temporal relationship graph by integrating a hybrid metric of cosine similarity and Dynamic Time Warping (DTW), and enhances the feature extraction by incorporating a network feature extraction module. This effectively addresses the limitations of traditional methods in feature representation and similarity measurement. In both simulated datasets and real-world scenarios such as the stock price data, the proposed method demonstrates good performance on classification metrics, with most classification accuracies stabilizing above 83%, and some categories approaching 90%. This validates the advantages of graph neural networks in capturing the spatiotemporal dependencies of time series, providing a new technical pathway for time series analysis.