Probabilistic distance based abnormal pattern detection in uncertain series data

Abstract

Abnormal pattern detection is an important task in series data anomaly detection. Because of the noise interference, the accuracy of abnormal detection method based on deterministic value is decreased. Whereas, most recent studies aimed at solving the anomaly detection problem in uncertain series use possible world models to describe the uncertainty in discrete data and select outliers as the anomaly detection objects. Abnormal pattern detection problems in continuous uncertain data are rarely reported.In order to improve the accuracy of abnormal pattern detection for uncertain data, we propose a Probabilistic Distance based approach for mining Abnormal Pattern Detection from uncertain series data (PD_ APD). Our considered approach re-express the Euclidean distance according to data’s probability density function (PDF), and get a probabilistic metric to compute the dissimilarity of two uncertain series. Our experiments show that, compared with Tarzan, a deterministic approach that directly processes data without considering uncertainty, PD_ APD provides a flexible trade-off between false alarms and miss ratios by controlling a probabilistic abnormality threshold. Especially, when data uncertain variance is large, PD_ APD has lower false alarms under the same specific miss ratio.