Research On Several Techniques In Time Series Data Mining

With the great development of information technology and the coming of the age of big data, how to deal with mass data efficiently becomes a research hotspot in cyber domains. Data mining technology has been receiving considerable scholars’attentions as well. Time series data mining is a significant research area in data mining field, which shows significant application value in many fields, such as the field of science, economy, meteorology, medical science and so on.In this thesis, first, some concepts and theoretical knowledge of time series data mining were introduced. Then research on current status and defects of time series data mining were described. The main content is consisted of three aspects, including research on similarity measurement, pattern discovery, forecasting in time series.First, to measure the similarity between time series more reasonably and more efficiently, based on point distance measurement and pattern distance measurement, a SAX-based composite metric method for time series similarity measurement was proposed. The experimental results show that the composite method has higher time efficiency and a higher degree of differentiation.Second, for research on pattern discovery, we synthesized both efficiency and quality of the algorithm, an improved algorithm for time-series pattern discovery was proposed based on key-point based segmentation algorithm and dynamic time warping distance. The experimental results prove that the proposed algorithm can measure the similarity between time series considering the existence of noise more accurately. And it is more robust under the situation of time series deformation.Last, for research on forecasting, in order to break the limitation of over dependence on historical data in traditional methods, an interval-similarity based fuzzy time series forecasting algorithm was proposed based on fuzzy theory. By forecasting the changing trend of future data instead of values, the algorithm obviously broaden the scope of application. The experimental results demonstrate that the algorithm is superior to other algorithms on forecasting accuracy.