Research On Incremental Time Series Classification Algorithm

Data mining technique has applied to a vast array of research areas in recent years, and types of mining data are more and more complex. Time series as one of popular issues has gained increasingly attention, such as stock trading, medical EEG, economic sales forecast, handwriting, human posture analysis, and so on. All above-mentioned data has common features that it is related orderly and real valued variables at the same time interval, and then such data with above characters is called time series. The following conclusion is that common method of data mining method can’t suit to time series data mining. And with the continuous development of the theory of big data and the updated changes constantly of data characteristics, incremental learning is essential in order to reduce temporal and space demand for implement of time series.This paper focuses on research of time series classification according to time series features of high dimensionality, ordered real-valued variables, autocorrelation and so on. The relevant research involves to time-series preprocessing, representation techniques, methods of distance measure and few aspects on time series classification, and summarizes the current state-of-the-art research status of time series classification method, such as shapelet based on decision tree for time series classification and shapelets transform for time series classification.First of all, from the perspective of image processing, this paper proposes a method for ITTS to transform image information to time series. As plant image information, the handwriting, and body posture data, the information that the naked eye intuitively access cannot be directly applied to time series classification method. So ITTS method in this paper can be obtained required time series data from the graphics information.Secondly, this paper puts forward a time series classification algorithm based on the incremental decision tree which is called ISDTC algorithm. The traditional time series classification algorithm can only handle the static data set, but advanced algorithm in this paper is able to handle the data set incrementally. The ISDTC algorithm is the process of time series classification based on the incremental decision tree. Experiments show that the eventually building decision tree is very similar by proposed ISDTC algorithm with the usage of static data.And lastly, this paper presents time series classification algorithm based on dynamically finding shapelets which is defined IPST algorithm. As the best approach to show a subsequence of time series, shapelets would be change dynamically as time goes on. Considering the thought, this paper proposes IPST algorithm that dynamically discovery current optimal k shapelets well, so as to improve the accuracy of time series classification.