Similarity Measure And Periodic Pattern Mining Of Time Series Data Mining

Data Mining has attracted much attention with the development of information technology. Time series data are a kind of important data existing in a lot of fields, such as financial market, industrial process, science experiments, etc, and the quantity of time series data has explosively increas. So it is necessary to study on the subject of the time series data mining. Nowadays, time series mining technology is still in its infancy and the algorithms are expected to be extended and to be complemented.After major issues in time series data mining surveyed, some algorithms are summarized and appraised in this dissertation. Then the problems on time series' representation, similarity measure and periodic pattern mining are deeply researched. At last, on the basis of the summery of the whole dissertation, we propose the several problems needed to be further researched in the future. The main innovative achievements are described as follows.1) A novel method of shape-based time series similarity measure is proposed. Based on the PLR algorithm, the time series is represented by using the relative change of the slope of the lines, which effectively reflects the degree of the dynamic change of the tendency of the curves. Moreover a corresponding distance measure formula is proposed, which can increase the robust to disturbances compared with the point-to-point Euclidean distance measure. The experimental results show that the approach can effectively measure the similarity of time series under various analyzing frequency.2) A novel method of local segmented dynamic time warping distance measure (LSDTW) for time series data mining is proposed. DTW is of great importance in time series similarity measure, but its expensive computation limits its application in massive datasets. The LSDTW algorithm is proposed to solve this problem. Based on the PLR algorithm, regarding every segment as a whole, classical DTW algorithm is used, and a compensate coefficient is proposed to ensure the accuracy of the method. The experimental results show that the approach gives better computational efficiency compared with the classical DTW without loss of accuracy.3) A highly efficient of asynchronous periodic 1-patterns mining method is proposed. A binary representation based mapping scheme is designed, and a modified dot product algorithm is proposed to find all the positions of an event in the time series; and a parallel calculation method is proposed to replace the series calculation method, which notably decrease the times of the calculation and access process. The experimental results show that the approach significantly increases the efficiency without loss of the accuracy.4) A novel concept and its mining method of local periodic frequent patterns in time series are proposed. Different from all exiting algorithms, this method not only can mine the frequent periodic patterns arised through the whole time series, but also can find the periodic patterns happened just in the locality of the sequence. This method first divides the time series into a set of fragments, then finds the hidden potential periodicities based on the data, and produces the local periodic frequent 1-pattern, at last using the max-subpattern tree method merges and outputs complex local periodic frequent patterns. The experimental results show that the approach can effectivly mine the local periodic frequent patterns in time series, and the cut algorithm and the periodicity threshold formula proposed in this paper can significantly increase the efficiency.