一种海量实时序列数据指数平滑预测分割算法*

时间序列分割是时间序列挖掘的重要任务之一。实时数据快速变化,数据量巨大,所以如何对实时数据进行快速而准确的分割很具有挑战性。本文提出基于指数平滑预测的滑动时间窗分割算法可以快速有效的分割在线实时数据,该算法基于滑动窗口和平滑指数算法,分析实时数据的统计特性,推导出序列的预测误差和压缩率之间的关系,通过序列预测的误差来判断分割点。加入校验环节提高算法的健壮性。通过本课题所使用的数据集以及公共数据集验证算法结果说明,该算法能够有效地在线检测出实时数据的分割点,并且时间复杂度较低。

One kind of massive real-time series data segmentation algorithm based on exponential smoothing prediction

Time Series Segmentation has been one of the most vital tasks in Time Series Mining. It is challenging to segment real-time series data quickly and effectively for the rapid changes in real-time data, and the huge amount of data. My paper proposes Sliding Window based on Exponential Smoothing Segmentation Algorithm, SWESSA, that can solve the problem. This algorithm based on Sliding Window and Exponential Smoothing Algorithm analyze real-time statistical properties, derive the relationship between sequence prediction error and the compression ratio make use of the error of series prediction to judge the split point. In order to improve the algorithm robustness, I add a verification section. The validation result of datasets used in this study as well as public data sets illustrate that the algorithm can effectively detect the split point in real-time data and has lower complexity.