Research On Time Series Data Mining

As a new kind of data analysis technique data mining develops fast. Many kinds of datasets can be the objects of data mining. Because time series are very common in datasets Time Series Data Mining (TSDM) has been one of the focuses of current data mining research.However,most TSDM literature focuses on similarity research and few on pattern discovery or rule discovery. What's more,little attention is paid to the complexity of time series in the field and the uncertainty of TSDM process is still to be explored.This dissertation researches on a new kind of TSDM frame which,taking into the inherent being of time series and combined with time series analysis techniques in the mining process,can extracting useful underlying system rules from time series. And then the rules can be used to analysis and predict the trend of the future. Moreover,fuzzy sets theory is adopted in the dissertation to deal with the uncertainty of the mining process and a new fuzzy frame of TSDM is given then.The main contents of the dissertation are summarized as follows according to the logical hieararchy:1. Based on Taken's Theorem,a new TSDM frame for univariate time series named States Evolution Patterns Mining (SEPM) is proposed. A series of new concepts are given and the mining process is investigated in details. Then the SEPM frame is extended for multivariate time series.2. A new concept of efficient support is proposed and its two liminal theorems are given and proved. The amelioration algorithm of fuzzy association rules discovery is raised based on efficient support. The uncertainty of the SEPM process is noticed. A fuzzy states evolution patterns mining (FSEPM) frame is brought forward. The implemetion process is investigated and then is ameliorated with efficient support.3. Key algorithms are tested and verified. Parameters impacts on the performance and results of mining parameters are experimented and analyzed. The performance of FSEPM and SEPM are compared.