Network Based Method for Time Series Clustering

Time series clustering is a prominent research area with broad applications. In light of the complexity and diversity of modern time series data, we propose a network-based method for time series clustering analysis. Firstly, we introduce a novel approach to construct networks from time series by segmenting them into non-overlapping equal-length segments and considering the correlations between them. This improved method effectively captures the potential patterns within the time series while reducing the node count compared to the overlapping segmentation method. Secondly, we develop a new clustering technique based on spectral theory of graph for analyzing time series data. By separately considering the correlation of both overlapping and non-overlapping segments, we transform the time series data into networks. Subsequently, inter-network dissimilarity measure on spectrum along with hierarchical clustering algorithm are sequentially applied to identify distinct clusters. To evaluate the performance of our method, we conduct simulation studies involving stationary and nonstationary, linear and nonlinear time series, as well as case studies using real-world datasets. The results demonstrate the effectiveness and robustness of our network-based clustering approach. The methodology exhibits the capability in handling highly correlated and nonlinear data, as well as discerning both trend and seasonal components.