Research On Time Series Classification Based On Discriminative Shapelets Extraction

With the explosive growth of time series data,time series classification has become an important research topic in time series data mining.Due to the shortcomings of time and space consumption and lack of interpretability of global feature-based time series classification,shaplets-based time series classification methods using local features,which are superior in terms of classification efficiency,robustness and interpretability,have attracted much attention in the field of data mining.The existing time series classification methods for extracting shapelets can improve the operational efficiency to a certain extent,but there are problems such as insufficient discriminative power,large-scale shapelets candidates and with inadequate interpretability.This paper mainly focuses on the extraction of discriminative shapelets in time series classification.To solve the above-mentioned problems,the corresponding solutions are proposed,and a large number of comparative experiments of the proposed methods are carried out on 28 public univariate data sets.The main research work and innovations of this paper are as follows:1.In view of the insufficient discriminative power of shapelets learned by existing models,the number of shapelets candidates is still large,and the lack of interpretability of the model,this paper proposes a two-phase filtering of discriminative shapelets learning algorithm for time series classification(TPSL-SGL).The sparse group lasso is introduced into shapelets learning model as a regularizer,and the time series are grouped by combining extreme points and sequence trends.The regularizer and local linear discriminant analysis are used to learn the location of shapelets.Then a twophase filtering framework is proposed to measure the sparsity of groups,so as to quickly locate the key group that plays a decisive role in classification.Finally,only the key group is used to extract shapelets for time series classification.The experimental results show that compared with the existing extraction methods based on shapelets,TPSL-SGL can not only improve the classification accuracy and time efficiency,but also reduce the scale of shapelets to some extent.Its classification accuracy is still competitive than several state-of-the-art non-shapelets methods.2.The TPSL-SGL model has difficulty in selecting appropriate sequences to participate in grouping,ignoring the effect of different grouping lengths on model construction and the lack of performance of the regularizer to ensure block sparsity of adjacent variables,this paper optimizes TPSL-SGL and proposes a shapelets extraction algorithm based on optimization and 2-phase filtering(SE2PF).SE2 PF has the following improvements over TPSL-SGL: finding standard time series to participate in grouping;the use of weight coefficients in the regularizer;the fusion penalty of fused lasso is introduced to optimize the regularizer.Those sparse regularization terms are combined as constraints to construct the objective function together with the local fisher discriminant analysis.And then,the two-phase filtering framework is used to accelerate the positioning of key group.Finally,only one key group is used to extract shapelets,train SVM and complete the subsequent classification tasks.The experimental results show that compared with some existing shapelets extraction methods,SE2 PF can significantly improve the classification accuracy,achieve competitive time cost,and reduce the scale of shapelets to a certain extent.Its classification accuracy also has advantages over several advanced non-shapelets classification methods.