Research On Time Series Representation Based Retrieval And Classification

Time series not only has the big data characteristics of "large volume,highdimensional",but also been incessantly accumulated,recognized as streaming times series.How to discover objective laws and potential knowledge within streaming time series has attracted widespread attention from academia and industry.In this dissertation,we present three in-depth studies on time series representation,retrieval,and classification.More details are listed as follows.Firstly,we propose an adaptive diverse time series representation method,called continuous segmentation and diversified rerepsentation(CSDR).CSDR could produce diverse representation results:Piecewise Linear Representation(PLR)and Symbolic Aggregate approXimation(SAX),while ensuring the relatively ideal representation efficiency.We have carried out corresponding comparative experiments between our CSDR method and baselines.By comparing the experimental results,our proposed method not only has high representation efficiency,but also can be used for multiple time series data mining tasks(visualization,clustering).Secondly,we present an efficient time series retrieval method based on multiresolution representation.This method can not only effectively represent the given time series based on multi-resolution representation criteria,but also effectively filter the candidate sequences that are not similar to the pattern sequences through index pruning strategy,thus improving the retrieval efficiency.More importantly,the adaptive pruning strategy established by our method satisfies the constraint of the lower bounding,namely,it can ensure the no-omission retrieval of the target sequence and achieve high-quality time series retrieval.Thirdly,we develop a temporal representation learning based time series classification model.It can not only efficiently obtain representative Shapelet sequences based on Tuning Points(TPs),but also realize the comprehensive temporal representation based on the enhanced bidirectional long short time memory network,thus effectively improving the classification performance.Moreover,this model can implement effective classification interpretability via temporal attention mechanism.