Research And Application Of Multivariate Time Series Association Mining Algorithm

With the development of computer technology, a large amount of historical data has been accumulated in various industries, among which, the time series data, as an important data type, is attracting more and more attention. Data mining, which is been widely used, has become one of the main tools for business institutions to mine the latent information in the data to create greater values. In the domain of network activities, medical diagnosis, and the weather forecast, time series data is the main object that been analyzed. Studying how to mine valuable information from time-series data has become increasingly practical significance.In this thesis, we focus on mining the association rules among multiple asynchronous time series. By means of frequent pattern mining and association rules analysis for multivariate time series, we can find out the hidden specific trends and special relationship in the time series, which including more information compared to other data. We investigated the existing association rules mining algorithms applied to time-series data, and analyzed their algorithm principles and the existing problems. Based on the temporal characteristic, we design a new framework for the multivariate time series association mining process.Our work mainly divided into two parts: preprocessing of time series data and proposing a new multivariate time series association rules mining algorithms based on time constraints. The preprocessing section consists mainly of four parts: defining the concepts of the time sequence and its segment model, obtaining the segment model sequences by data fitting on original time series using a bottom-up line fitting method, filtering the anomaly pattern via a new cleaning method based on the density peak, and clustering the segment models utilizing clustering algorithms that are based on density peak. By building the frequent pattern tree, the new algorithm implements the association analysis for multiple time series based on time constraints and takes into account the time delay property.In the end, we conducted our experiments on multiple data sets, and analyzed the results of association rules for stock data set via clustering algorithms based on density peak and via association mining algorithms based on time constraints.