Multivariate Time Series Clustering Algorithm Based on Improved K-Nearest Neighbor Network

To address the limitations of existing clustering algorithms for multivariate time series, such as weak spontaneity and poor clustering performance, this paper proposes a multivariate time series clustering algorithm based on an improved K-Nearest Neighbor (KNN) network. By introducing the shape-based weighted dynamic time warping similarity measure, the similarity between multivariate time series objects is calculated. Subsequently, a directed weighted KNN network is constructed, where nodes represent multivariate time series objects, edges denote KNN relationships, and weights correspond to similarity values. Based on this network model, the K-Means algorithm is used to determine the cluster centers, and the Blondel-Guillaume-Lambiotte-Lefebvre hierarchical community detection algorithm is employed to cluster the time series objects. Numerical experiments demonstrate that the proposed method outperforms other related clustering method in terms of accuracy and robustness. Sensitivity analysis of the value of K in KNN clustering further validates the effectiveness of the parameter selection strategy.