基于集成预测的稀有时间序列检测

Outlier detection in time series through neural networks forecasting

为了解决误判问题,从预测的角度给出了离群点的定义,并提出了预测可信度和离群度的概念;同时,提出采用置换技术来降低离群点对预测模型的影响,并提出了基于集成预测的稀有时间序列检测算法。针对真实数据集的实验表明,可信度和离群度的定义是合理的,稀有时间序列检测算法是有效的。

From the view of forecasting, a novel definition of outlier in time series was presented, as well as the definition of the forecasting confidence and the degree of outlier. The technique of permutation was proposed to alleviate the impact of outliers upon the forecasting model. To solve the false alarm problem, the forecasting-based outlier detection algorithm was pres- ented. The experiments conducted on the real-world datasets show that definition of the degree of outlier is reasonable and the outlier detection algorithm is effective.