I have read a lot about the robust anomaly detection of Netflix which they open sourced as part of their Surus Project (https://github.com/Netflix/Surus). The project anomaly detector is based on the Robust PCA approach by Candes et al. (https://statweb.stanford.edu/~candes/papers/RobustPCA.pdf) where the data matrix is resolved into a low rank matrix and a sparse matrix. There are different approaches to get this decomposition like Principal Pursuit etc., I am unable to locate anywhere how this RAD algorithm is utilizing the RPCA. I am trying to recreate it for some of my own purpose as an anomaly detector for a time series, and i am able to reach upto the decomposition step but am clueless as to what to do next.
Can anyone please help me as to how the anomalous values are detected? Is the sparse matrix utilised. Any help on this topic would be helpful. I am working with a time series so any help along those lines would be wonderful although that is not a restriction. Any general idea on how to designate a point as an anomaly out of the given data matrix would be useful. Thanks.
