Study On Time Series Classifier And Sales Forecasting Based On Improved Time Series Algorithms

In the information age of 21st century,the time-series data generated by people's production and life grew rapidly.Mining potential,valuable rules,patterns,and knowledge from massive,continuous,high-dimensional,and noisy time-series data,and then build mathematical models for analysis,description,prediction,and application had become an effective method in deeply utilizing massive data resources.The research hotspots for time-series data mining include:model representation,similarity measurement,classification and clustering,and prediction.Almost all the time-series data mining tasks are requires a basic notion of similarity measures.A advanced and efficient measurement method has critical effect on the performance of subsequent time series data mining.The dissertation conducts effective research on time series similarity measurement and predictive analysis methods.Since Traditional Dynamic Time Warping(DTW)measurement method is prone to over-bending and has the shortcoming of high computation complexity and low efficiency,an Updating Dynamic Time Warping(UDTW)measurement method based on path correction was proposed.The method can effectively extract the sequence feature information through a segmented dimensionality reduction method,set dynamic penalty coefficients for over-curved paths,correct the path,s degree of curvature,and reduce the computational cost of metric distances.An auto-regressive moving average model(ARMA)was constructed to predict and analyze the sequential sales data of retail commodities and forecast the trend of the sales volume of commodities.The Support Vector Regression(SVR)method and Extreme Learning Machine(ELM)method are used to predict and compensate the nonlinear errors of ARMA time series model to enhance the accuracy of prediction.And the paper proposed a novel selective ensemble method to determine the parameter of ARMA which also simplified the process of ARMA.Based on updating distance metric,1-nearest neighbor classification algorithm was used to classify time series data.Experiments over different datasets confirmed that the algorithm is superior to the original DTW and other variant methods,which improved the execution efficiency without decreasing classification accuracy.The article collects sales data to compare and analyze the performance of the combined forecasting model,and proves that the proposed method has good robustness and is insensitive to noise data.