基于多元时间序列的自适应贪婪高斯分段算法

现有多元时间序列分段算法中分段点的选择以及分段个数的确定往往需要分别独立完成,大大增加了算法的计算复杂度.为解决上述问题,提出一种基于多元时间序列的自适应贪婪高斯分段算法.该算法将多元时间序列各个分段所对应的数据解释为来自不同多元高斯分布的独立样本,进而将分段问题转化为协方差正则化的最大似然估计问题进行求解.为提高学习效率,采用贪婪搜寻方法使每个段的似然值最大化进而近似地找到最优分段点,并且在搜寻的过程中利用信息增益方法自适应地获取最优的分段个数,避免分段个数确定和分段点选择分别独立进行,从而减少计算的复杂度.基于多种领域的真实数据集实验结果表明,所提出方法的分段精度以及运行效率均优于传统方法,并且能够有效完成多元时间序列的异常检测任务.

Adaptive greedy Gaussian segmentation algorithm based on multivariate time series

For most multivariate time series segmentation algorithms, the selection of segmentation points and the determination of the number of segments often need to be completed independently, which greatly increase the computational complexity of the algorithm. In order to solve the above problem, an adaptive greedy Gaussian segmentation algorithm based on multivariate time series is proposed. The algorithm interprets the data points from the segmentations of multivariate time series as independent samples of different multivariate Gaussian distributions, and then transforms the segmentation problem into a covariance-regularized likelihood maximization problem to solve. In order to improve the learning efficiency, the greedy search method is adopted to maximize the likelihood value of each segment to find the optimal segment point approximately. During the search process, the information gain method is adopted to adaptively obtain the optimal number of segments, which avoids from realizing the determination of the number of segments and the selection of segment points independently to reduce the computational complexity. The experimental analysis is carried out on real datasets in many different fields. Compared with traditional methods, the proposed method can obtain higher accuracy and efficiency, and is able to detect outliers in multivariate time series effectively.