Study Of Linear Representation Method And Similarity Measurement Algorithm Of Time Series

Time series data is an important complex data, existing in natural phenomenon and social economy. Data mining is applied to analyze time series data had found information contained in time series, so, the research on data mining of time series is one of important hot spots of data mining. The study on data mining of time series mainly concentrates on the linear representation and the similarity search.Before undertaking operation, time series must be preprocessed, to receive high quality data. Linear representation of time series is a basic approach of data preprocessing. This paper researched existing linear representation algorithms, summed up their defect, presented a new linear representation method-trend variation point. Run mode of time series is called "trend", the point that the trend before the point and the trend after the point is different is called trend variation points. Trend variation point algorithm was based on the idea of "trend", noise was immunized, and trend variation points was found in monotonic sequence.In this paper, research on the existing similarity measurement for time series data was undertaken, summed up the defect about scaling on time axis. To resolve this defect, this paper use the idea of "model-distance" from model-distance algorithm, and "trend model" was raised and restrict of time span was into algorithm, then trend model distance measure algorithm based on trend variation point was raised. This algorithm not only resolve the defect about time axis, but also forecast future trends for time series which was an innovation in similarity measurement, resolving the defects of existing similarity measurement effectively.Forecast of agricultural production is a significant element task for establishing agriculture policy, developing agricultural production and agricultural trade. In this paper, above-mentioned research achievement was applied into time series of agricultural production, result of the first experiment verified that Trend variation point algorithm could find trend variation in time series more effectively, providing high-quality sequence presentation for further action; result of the second experiment verified that trend model distance measure algorithm possesses higher veracity in application, and it is feasible that merged forecast into similarity measurement.