oddstream (version 0.5.0)
Outlier Detection in Data Streams
Description
We proposes a framework that provides real time support for early detection of
anomalous series within a large collection of streaming time series data. By definition, anomalies
are rare in comparison to a system's typical behaviour. We define an anomaly as an observation that
is very unlikely given the forecast distribution. The algorithm first forecasts a boundary for the
system's typical behaviour using a representative sample of the typical behaviour of the system. An
approach based on extreme value theory is used for this boundary prediction process. Then a sliding
window is used to test for anomalous series within the newly arrived collection of series. Feature
based representation of time series is used as the input to the model. To cope with concept drift,
the forecast boundary for the system's typical behaviour is updated periodically. More details
regarding the algorithm can be found in Talagala, P. D., Hyndman, R. J., Smith-Miles, K., et al.
(2019) .