Dissertation > Excellent graduate degree dissertation topics show

Research and Application of Cluster Algorithm in Time Series

Author: HanNa
Tutor: ZuoShaoHua
School: Guangdong University of Technology
Course: Computer Software and Theory
Keywords: Time Series EMD Hadamard transform Matrix similarity measure K-means algorithm
CLC: O211.61
Type: Master's thesis
Year: 2011
Downloads: 149
Quote: 0
Read: Download Dissertation

Abstract

Time series data is ubiquitous, and cluster analysis of time series can get a lot of valuable time-related information hidden in the time series data to achieve the acquisition of knowledge, knowledge under the guidance of the activities. However, the very large amount of time-series data in real life, therefore, cluster analysis of the time series before its dimension reduction. The main work of this paper: (1) A weighted value matrix similarity measurement method. Multivariate time series using singular value decomposition method to get the feature matrix and singular value vector, respectively, as the matrix and the right value. Hadamard transform feature matrix and the correlation coefficient matrix eigenvalue as the matrix and the right value. (2) gives a multivariate time series clustering analysis method based on EMD and SVD. The original data to fill a vacancy values ??normalized after the first use of EMD multivariate time series trend extraction smoothed sequence, then using SVD multivariate time series length dimension reduction last of the characteristic matrix of the multivariate time series and the corresponding weights to improve the K-means algorithm clustering analysis. (3) gives the multivariate time series clustering analysis method based on Hadamard transform. Carried out the filled values ??as well as standardized raw data, the first use of the Hadamard transform dimensionality reduction of multivariate time series, while taking advantage of the wavelet transform sequence dimensionality reduction. Then last for pretreatment data to determine the different length feature matrix corresponding weights, improved K-means algorithm clustering analysis. This paper describes the EMD and SVD multivariate time series clustering methods and multivariate time series clustering method based on Hadamard transform. Both methods have their advantages: First method First, the time series of EMD, extraction time sequence trend. EMD filtering of the noise in the time series generated trend sequence able to accurately reflect the original sequence of the trend toward making the sequence can become clearer, and the information is lost relatively small, and so reduce the dimension on the basis of the trend sequence , and to improve the clustering effect. Second, the trend series SVD decomposition sequence of different lengths can be unified to the same scale. Usually most of the cases the time series of unequal length, the dimension of the sequence features extracted by SVD features only parameter, regardless of the length of the sequence. This makes the the unequal length sequence clustering possible; second method, first, by the Hadamard transform sequences dimensionality reduction. Because the Hadamard transform sequence data after the energy concentration is relatively high, it can be a very short data indicates that the original sequence and to maintain the original sequence trend transform morphology. Cluster analysis based on this sequence dimensionality reduction can largely improve clustering accuracy required time and clustering. Proved by experiments, both clustering methods to achieve effective clustering of multivariate time series, the two parallel clustering algorithms are analyzed and compared.

Related Dissertations

CLC: > Mathematical sciences and chemical > Mathematics > Probability Theory and Mathematical Statistics > Theory of probability ( probability theory, probability theory ) > Random process > Smooth process and the process of second-order moment
© 2012 www.DissertationTopic.Net Mobile