Research On Dynamic Fuzzy Clustering For Time Series

With the rapid development of information technology, the massive data occur in variousfields. It is one of the important topics how to discover the valuable knowledge from the hugedata in computer science and other application fields. Data mining comes into the world toprovide a powerful tool to deal with this problem. Data mining also goes by the name ofknowledge discovery in database (KDD), aiming at searching for the useful knowledgehidden in massive data. As an interdiscipline with promising future, data mining integrates themature tools and techniques of many subjects, such as database technology, statistics,machine learning, pattern recognition, artificial intelligence and etc. The main topicsdiscussed in data mining include association rules, classification, clustering, prediction,sequence found, anomaly detection and so on.In the past decades, the research on time series has aroused tremendous interest amongmany fields, such as finance, commerce, meteorology and etc. Meanwhile, its complexity liesin high dimension, noise, stretching and translation on extent, expansion and contraction onthe time axis, linear drift and discontinuous points. These properties of time series lead to achallenging problem while clustering them. The most of existed methods to cluster time seriesare a direct application of some approaches which are usually employed to cluster static data.The time property of time series is often neglected in these methods. Therefore, it isworthwhile to research on dynamic fuzzy clustering time series.This dissertation discusses the fuzzy clustering validity and the dynamic clustering oftime series. The results are briefly described as follows.1. Based on fuzzy c-means algorithm, a new fuzzy clustering validity index is presentedby computing the compactness within clusters and overlap between groups via membershipdegree matrix. The optimal cluster number can be effectively found by the proposed index forthe data set which cluster results are overlapped each other. Generally speaking, a goodpartition will result in that the compactness within the class is strong and the overlap betweenclasses is small. The proposal overcomes one of the shortcomings of FCM, i.e. pre-assign thecluster number c. The experimental results show the validity of the proposed index. It is alsocan be seen that the optimal cluster number are obtained for three different values m=1.8,2.0and2.2generally used in FCM algorithm.2. To overcome the shortcomings of using static clustering algorithm to group time series,a dynamic clustering algorithm for time series is introduced. The algorithm firstly uses the set of key points to represent the time series, in order to achieve the purpose of reducingdimension. And then time series is partitioned by the improved FCM algorithm. Finally thecluster label of switching series will be found over time. The proposal reveals that the clusterlabel of some series will be changed over time. This is exactly the difference between staticdata and time series when clustering them. The traditional FCM algorithm is improvedthrough L-W distance, which is not sensitive to singular value. The experimental results showthe method proposed in the paper have the feasibility and validity.