基于两点的时间序列相似性研究

Research on the Similarity in Time Series Data with Two Points

目前,时间序列相似性判定大多采用欧式距离和动态时间弯曲DTW(Dynamic Time Warping)方法,这两种方法均存在一定缺陷。欧式距离要求序列长度一样,垂直移动序列将影响相似性判定和阈值设置的经验性;动态弯曲距离对欧式距离进行了优化,避免了欧式长度的一致性,但其他两个缺点仍然存在且计算复杂度增加。提出了一种新的基于两点时间序列相似性算法,可计算任意两序列的相似度。首先分析了两点组成的序列形态,提出了相似性判定方法TPSS(Two Points Segmentation Similarity);其次为提高相似性判定的鲁棒性,减少人为阈值设置的影响,对TPSS进行了拓展;最后给出了算法及实验分析。实验结果表明,该算法能很好地判定任意序列的相似性,提高了鲁棒性及减少人为干预,对数据挖掘中的聚类与预测有很好的帮助作用。

At present, the judgement of the similarity of time series is based on Euclidean distance and DTW （ Dynamic Time Warping）. These two kinds of methods have some defects. Euclidean distance request the same length of sequence, vertical movement will affect similarity judgment of sequence. Anyone set the threshold of the empirical is difference ; DTW is optimized of the Euclidean distance to avoid the consistency of European length. But the other two disadvantages still exist and the computa- tional complexity is increased. This paper proposes a novel similarity algorithm of time series based on the two points, it can cal- culate any of the two sequence similarity. First, it analyzes the time series form of two points, put forward TPSS method. Second- ly, in order to improve the robustness of the similar judgment, reduce the influence of artificial threshold setting, TPSS is devel- oped. Finally, it gives the algorithm and experiment analysis. The experimental results show that tThis algorithm can effectively determine the similarity of arbitrary sequences, improves the robustness and reduces human intervention and can help clustering, prediction in data mining.