Abstract: This paper presented a DTW distance-based sampling framework to effectively improve the accuracy of cyclic time series prediction in large-scale datasets.It addresses the problem of noisy identification for each given prediction condition,and formalizes the impact of noise with the SVR-based predicting method.On top of the DTW-based similarity measurement,this paper presented an end-to-end identification method to improve the quality of the training set.It also introduced a regularized function in the kernel function of SVR,such that the generalization error can be minimized based on the distances between each training instance and the prediction condition.The experiment conducts a series of widely adopted cyclic time series to evaluate the precision and stability of the proposed method.The results demonstrate that in terms of high-quality training instances and the weighted regularization strategy,the proposed method remarkably outperforms its competitors in most of the datasets.

Key words: Data mining, DTW, Prediction, Similarity, Support vector machine, Time series