时间序列数据的稳健最优分割方法

针对分段多项式回归方法存在计算效率低和对噪声较敏感等缺点，提出了具有稳健性的最优分割方法，以解决时间序列数据相似搜索及知识发现处理中的长时间序列分割问题。该方法采用自顶向下策略，然后根据自适应定阶算法直接选定一个合适的多项式阶，对每个候选变化点，经过一次判断即可确定多项式的合适阶次。由于该方法基于对线性模型的数据矩阵作奇异值分解，从而可自适应确定子序列合适的模型，简化了计算过程，文中对此给出了理论证明。通过与Garalnik-Srivastava方法进行实验比较，证明所提方法不仅计算效率高，而且具有良好的稳健性。

Robust Optimization Segment for Time Series Data

Aimed at some shortages in the existing methods based on piecewise polynomial representation, such as sensitive to noise and lower computational efficiency, a robust method of optimal segmenting time series is proposed. The suggested method can be used to segment large time series into some subsequences to solve this problem in similarity search and knowledge discovery. Up to bottom policy is adopted, and an algorithm of adaptive determining orders of polynomial functions is developed. At each candidate change point, a reasonable order of the polynomial function can be determined by once estimating.The orders of polynomial functions are selected adaptively by means of singular value decomposition of data matrix of piecewise polynomial models, so the computing process is simplified, and theoretical proof is presented. Compared with Guralnik Srivastava method, the experiment results demonstrate that proposed method not only reduces searching time significantly, but also has characteristics of stabilization and robustness.