Research On Dynamic Clustering For Time Series

With the advent of the big data age,data mining has become a hot issue.And time series clustering as an important part of data mining,has aroused tremendous attention.Because of the time series data with the properties of high dimension,massive volume and volatility over the time,then the traditional clustering algorithm could not describe its characteristics which changing with time.For this problem,this paper proposes a dynamic clustering algorithm for time series based on fuzzy clustering.In this paper,the time series is reduced by extracting the key points from time series,and makes the key sequence equal by combining the key sequence's subscripts.This way not only maintains the effective information in the data,but also greatly reduces the dimension of time series and the data storage space of data.And for the time series with the important feature of changing over time,the traditional FCM algorithm is improved.Considering to mixing the Lange distance,the variance distance,the kurtosis distance and the skew distance with reasonable weights to replace the European distance,then measuring the similarity between time series.This method not only avoid the influences of the singular value in the sequence,but also could measure the similarity between the time series with translation or stretching,which greatly improves the accuracy of the clustering algorithm.In addition,the paper proposes a dynamic clustering algorithm based on the improved FCM algorithm,which can cluster the fluctuating time series into different categories in different time periods.In order to evaluate the performance,feasibility and practical value of the algorithm,this paper uses the data set named phoneme and the data from lsolet5 database to test the the influence of the parameters in the improved FCM algorithm.Then,using the data set named Synthetic Control and the transformed data from lsolet5 database to compare the performance of traditional clustering algorithms with the improved FCM algorithm.Finally,the dynamic clustering algorithm is used on the data called beef and the real financial time series.The experimental results show that the proposed algorithm can describe dynamic characteristics of time series.