Unsupervised outlier detection for time series by entropy and dynamic time warping
Unsupervised outlier detection for time series by entropy and dynamic time warping
复制标题
DOI：
10.1007/s10115-017-1067-8
复制
发表时间：
2018-02-01
影响因子：
2.7
通讯作者：
Canitia, Bruno
中科院分区：
文献类型：
--
作者：
Benkabou, Seif-Eddine;Benabdeslem, Khalid;Canitia, Bruno
在过去的十年中，时间数据的离群值检测已从数据挖掘和机器学习社区引起了很多关注。尽管其他作品通过双向方法（相似性和聚类）解决了这个问题，但我们在本文中提出了一种同时处理这两种方法的嵌入式技术。我们根据熵和时间序列的动态时间扭曲将异常检测的任务重新制定为一个加权聚类问题。然后，通过适合此类数据的新提出的成本函数的优化问题检测到异常值。最后，我们提供了一些实验结果，以验证我们的建议并将其与其他检测方法进行比较。
In the last decade, outlier detection for temporal data has received much attention from data mining and machine learning communities. While other works have addressed this problem by two-way approaches (similarity and clustering), we propose in this paper an embedded technique dealing with both methods simultaneously. We reformulate the task of outlier detection as a weighted clustering problem based on entropy and dynamic time warping for time series. The outliers are then detected by an optimization problem of a new proposed cost function adapted to this kind of data. Finally, we provide some experimental results for validating our proposal and comparing it with other methods of detection.