[infolabe-anomalies] Updated anomaly detection algorithm

Joss Wright joss.wright at oii.ox.ac.uk
Tue Aug 16 09:54:45 UTC 2016


Hello All,

Just a note that the anomaly detector has now been updated to
pre-process the metrics data to remove weekly cycles. ("Seasonality" is
the official term).

For reference, we use Seasonal Trend Decomposition bia Loess Regression,
detailed here: http://www.jos.nu/Articles/abstract.asp?article=613 ,
using robust fitting due to the expected anomalies in the data. (R
manual page here:
https://stat.ethz.ch/R-manual/R-devel/library/stats/html/stl.html

This should make the detector more robust with respect to countries that
show significant regular weekly dips and spikes in usage. (Ethiopia,
Bangladesh, and Mongolia are all examples that regularly show up as
anomalous that also show weekly seasonality.)

Joss
-- 
Dr. Joss Wright | Research Fellow 
Oxford Internet Institute, University of Oxford
http://www.oii.ox.ac.uk/people/?id=176
Faculty Fellow, Alan Turing Institute


More information about the infolabe-anomalies mailing list