Research On Multivariate Time Series Data Classification Method Based On Metric Learning

Time series data is one of the types of data that people often encounter in their daily lives and work.It is widely used in medical,financial,industrial and meteorological fields.Compared to the single attribute of univariate time series data,it is often necessary to record multiple attributes at the same time at each moment,so the number analysis of multivariate time series data is closer to reality.Therefore,in-depth study of multivariate time series data analysis has more and more important practical significance and application value.From the basic characteristics of multivariate time series data,and deeply studies the similarity measure methods of existing multivariate time series data.The classification effect of the existing MD-DTW method is not very good.The similarity measurement method of time series data is deeply studied,and the classification method for multivariate time series data is proposed.Improvements for multivariate time series data classification are proposed in two research directions.In the first direction,the dynamic time warping algorithm based on Mahalanobis distance(MD-DTW)is a commonly used multivariate time series data similarity measure method.In order to calculate the local distance of time series data,only the spatial distance is considered.In this paper,the MD-DTW method(Sp-MD-DTW)with fusion Spearman coefficient is proposed.The algorithm uses the Spearman correlation coefficient and the Mahalanobis distance to construct a new local distance calculation formula,and obtains a better local distance metric,which improves the similarity measurement effect of the MD-DTW method.Then based on the proposed Sp-MD-DTW similarity measure method,the PGDM metric learning algorithm is merged,and the time series data classification method based on Sp-MD-DTW and PGDM metric learning algorithm is proposed.The method uses the PGDM metric learning method to calculate a Markov matrix suitable for the current task from the training sample set.The Markov matrix is used as the metric parameter of the similarity measure method,which can better represent the similarity between the training set and the test set.measure.In the second direction,the global distance measurement algorithm between the currently used time series data is generally based on dynamic time bending and only calculates the spatial distance.This paper proposes a time series data classification method based on K-S test.The method uses K-S test to describe the significant difference of time series data,constructs a new global distance metric formula with p value as the parameter,and completes the time series data classification by k-NN classification method.Finally,the proposed time series data classification algorithm based on the K-S test is combined with the proposed time series data similarity measurement method based on Sp-MD-DTW and PGDM to obtain a classification method for multivariate time series data.Finally,the published UCI dataset is used to verify the effect of the proposed algorithm.The experimental results show that the proposed algorithm achieves the desired results.