Evolution-based Clustering Algorithms For Time-Series Data And Their Applications

Generally speaking,data can be divided into static data and time-series data.Traditional clustering analysis is mainly for static data.However,in practical application,data always changes as time elapses,such as stock data,social media data,etc.Unlike the static data,it is necessary to cluster the data for every time step when processing time-series data.Therefore,it is required that the clustering method should not only reflect the long-term clustering trend,but also have robustness and certain smoothness to short-term changes.One important method of constructing those clustering algorithms is to add a penalty term to the cost function of the traditional static clustering methods by time smoothness^[1][2],which can be called as penalty methods.Under this framework,various static clustering algorithms,such as fuzzy clustering,K-means clustering,spectral clustering,etc.,can be conveniently extended to time-series data clustering algorithms.However,the shortcomings of these algorithms still exist,such as being sensitive to initial value and being easy to converge to local solutions,etc.In this paper,according to the existing drawbacks of these algorithms,we carry out the corresponding researches and improve some algorithms based on the characteristics of time-series data.The specific work is as follows:?1?Combining differential evolution with evolutionary spectral clustering algorithm.Constructing evolutionary spectral clustering method based on differential evolutionary algorithm which can achieve global optimization by random to overcome the drawback of local convergence.?2?Proposing an evolutionary fuzzy clustering algorithm for time-series data.According to the smoothness requirement in time series,the paper uses the weighted method to deduce the evolutionary fuzzy clustering method and therefore obtains two clustering schema.We also implement the algorithms by using differential evolutionary methods.?3?Correcting the distance metric of the sample data to get semi-supervised clustering of time-series data.The similarity matrix of the sample data is modified by using the auxiliary information positive association sets?ML?and the negative association sets?CL?.In this paper,two methods are used.One is to correct the distance metric between ML and CL sets based on the data density;the other is to learn a new distance metric from the time-series sample data through ML and CL sets.?4?Using the evolutionary spectral clustering algorithm based on differential evolution to group stock data obtained by the Web spider and to predict the trend of stock market through the corresponding clustering method.