Research On Time Series Classification Based On Discrete Cosine Transform

Time series data generally is data that change sequentially, these data generally measured inan equal time interval. How to manage and use these data efficiently, and how to mine the rulesand knowledge hidden in them attract wide interests in industry and academia. Time seriesclassification is one of the important tasks of time series data mining. Time series classificationuse the whole series as the input, and the aim of classification is giving this series a discretesymbol. Because time series have the feature of massive, high dimensionality, high noise, timeseries classification is more complex than general classification problem, and this makes thatgeneral classification algorithm can’t be applied to time series classification directly. In therecent ten to twenty years, researchers developed many algorithms to the classification of timeseries, based on these research achievements, this paper makes related research on thedimensionality reduction and classification of time series. The main work of this paper is thefollowing respects:1. Using discrete cosine transform to reduce the dimensionality of time series. Thedimensionality reduction of time series is the base of time series data mining. Because of thefeature of high dimensionality of time series，directly mining work on the original series willcause “dimensionality curse”, so generally, before the data mining on the time series, thedimensionality of time series should be reduced. Discrete Fourier transform, discrete wavelettransform, and piecewise linear approximation etc. have been developed in the practicalapplications for the series transform. All these methods have some limitations, so developingmore robust method has very important meaning. As a real number domain transform, discretecosine transform overcomes the shortcomings of the plural domain transform of discrete Fouriertransform. This paper uses discrete cosine transform to reduce the original series, and carries onthe similarity search and classification work, and achieves good effect.2. The classification of time series based on discrete cosine transformation. Thedimensionality reduction of time series using discrete cosine transform，makes the base of thenext classification work. When classify the time series, distance based classification method is a classical and common method. K-nearest neighbor algorithm is a mature method in theory, andas an improvement of k-nearest neighbor, weighted k-nearest neighbor algorithm overcomessome disadvantages of k-nearest neighbor algorithm, and the correct rate improves obviouslywhen using it to classify. This paper use k-nearest neighbor algorithm based on Euclideandistance to classify time series that processed by discrete cosine transform. The simulation showsthat this method has a well classification correct rate.The research result of this paper shows that using discrete cosine transform for thedimensionality reduction of time series can extract the features of series efficiently, andcombining weighted k-nearest neighbor algorithm to the further classification job has a goodperformance，and this makes the base for the further research work.