Research On Time Series Data Mining Based On Similarity Analysis

Data Mining has attracted a great deal of attention under the development of information technology. After surveyed major issues in data mining, problems on Time Series' Representation, Distance Measurement and Clustering were analyzed. Some algorithms and solutions were proposed. The main ideas in this dissertation are listed in two parts as follows:(1) In order to measure the dynamic trend and content similarity of time series effectively, Tuple Vector Representation of time series was redefined. Through projecting every endpoints of tuple sequence onto the other sequence, we contructed Trimness Tuple Sequence, defined Teimness Distance for sloving the time series whole matching problem.This distance overcomes the problem of time series mismatch based on point distance. Then, we proposed TVTW(Tuple Vector Time Warping) for sloving the time series scaling matching problem based on DTW(Daynamic Time Warping), which support tuple sequence stretching along time more reasonable and senseable comparing to stretching of DTW. And, the consequences of experiment are approving.(2) Applied TVTW technices to time series clustering, we proposed HTMC ( Hierarchical then K-means Clustering ) time series clustering framework under TVTW distance metric. The most important issue in clustering is creating central sequence on which we definded merging operation. The output of the operation is also tuple sequence; we called 'Temple Tuple Sequence', which megre sequences to combine information from sequences in same classes. We take advantage of temple tuple sequence as central sequence in each step of clustering. Due to Lowe-Bounding of TVTW technology we proposed in this section, some of candidates were pruned away from candidate set before actual execute time consuming TVTW compution, efficiency of our clustering algorithm was accelerated significantly.Finally, the dissertation is concluded with a summary and prospect of future work.