Title:Multi-scale streaming anomalies detection for time series

Abstract:In the class of streaming anomaly detection algorithms for univariate time series, the size of the sliding window over which various statistics are calculated is an important parameter. To address the anomalous variation in the scale of the pseudo-periodicity of time series, we define a streaming multi-scale anomaly score with a streaming PCA over a multi-scale lag-matrix. We define three methods of aggregation of the multi-scale anomaly scores. We evaluate their performance on Yahoo! and Numenta dataset for unsupervised anomaly detection benchmark. To the best of authors' knowledge, this is the first time a multi-scale streaming anomaly detection has been proposed and systematically studied.