On the Ranking of Variable Length Discords Through a Hybrid Outlier Detection Approach

In this paper we are interested in identifying insightful changes climate observations series, through outlier detection techniques. Discords outliers that cover a certain length instead of being single point the time series. The choice can be critical, leading to works on computing variable discords. This increases number discords, with potential overlapping, subsumption and reduced results. work introduce hybrid approach rank discords extract most prominent ones, yield more impactful We propose ranking function over extracted accounts for contained anomalies. investigate combination pattern wise anomaly detection, Matrix Profile paradigm, two different detectors. experimented MAD PROPHET algorithms based concepts tested our observations, representing monthly runoff series between 1902 2005 West African region. Experimental results indicate combined method, yields qualitative rankings, an extraction higher values extreme events within