Research On Multivariate Time Series Classification Based On Gaussian Graphical Model

Multivariate Time Series(MTS)classification problems broadly exist in numerous fields,including pattern recognition applications,medicine,multimedia and biological sciences.In the research of MTS classification,traditional methods need to measure the similarity between different sequences,or select and extract effective features from the original sequence data,or rely on deep learning networks for classification.These methods have some shortcomings,such as unable to deal with time series data with variable length,or present a high time complexity,or the classification results are not interpretable.Therefore,it is of great significance to deal with MTS with variable length,reduce the time complexity of the MTS classification process,and interpret the classification results by mining the complex dependencies between different variables.Based on the above exploration,this dissertation studies the problem of multivariate time series classification.The main contributions of this dissertation are as follows:1)In this dissertation,the Graphical Gaussian Model(GGM)is used to represent the MTS data,that is,to convert the MTS data into the GGM parameters(mean matrix and inverse covariance matrix).This representation provides a set of model parameters that do not depend on the length of MTS,so it can handle MTS with variable length.2)To highlight the complex dependencies between different variables,the Alternating Direction Method of Multipliers(ADMM)is applied to sparsely solve the important parameter of GGM--inverse covariance,and obtained the sparse inverse covariance parameter.In order to learn the sparse inverse covariance used to accurately describe the subsequence,a learning method based on Log Det divergence is established in this dissertation.The accurate and robust sparse inverse covariance representing the subsequences of each class in the MTS dataset is obtained through the Expectation-Maximization algorithm.3)Based on the established model parameter space,this dissertation proposes several different distance measurement methods to construct model classifier. Experiments conducted on several public MTS datasets and a case study demonstrate that our proposed approach performs better in the classification accuracy and interpretability compared to others well-known approaches.