Research On Feature Representation And Similarity Measurements Of Time Series

Research on feature representation and similarity measurements of time series is one of the hottest issues in the field of data mining. It is the foundation of time series classification and clustering and has a broad application prospect.Piecewise aggregation approximation (PAA) is one of the most commonly used methods of feature representation. PAA algorithm treats each section on average, which is one of the shortcomings of this algorithm,aiming at this shortcoming we present a method based on wavelet entropy and piecewise aggregation approximation in this paper, we treat wavelet energy entropy as an indicator of the complexity in one section. According to the wavelet entropy value of each section to allocate points within each section. We use matlab to deal with the five different types of data selected from UCI data set. The experimental results show that compared with PAA,the fitting error of PAA_WE is smaller, PAA_WE can be more accurate than PAA in representing time series.LB PAA lower function is a commonly used lower bound of the DTW distance, In this paper, we have proved that compared with PAA, PAA_WE can represent time series more accurately, so we introduce the PAA_WE method to the lower bound function area, and put forward a LB_PAAWE lower bound function. We compare the tight degree of four lower function through the matlab simulation experiment. The experimental results show that: LB PAAWE lower function compared with the other three lower function listed in this paper,is more closer to DTW distance,LB_PAAWE is a relatively tight lower function.This paper mainly studies the feature representation and similarity measurement of time series, the mainly work includes:we put forward a PAA_WE method in feature representation field and apply PAA_WE to lower bound function theory, and put forward the LB PAAWE lower function.