The Anomaly Detection survey by Chandola et al categorizes anomalies into point anomalies and collective anomalies. Do we need this type of categorization in anomaly detection in time series, though? Can't I always map time series data point of a time range into a single data point (e.g., a representative vector), and then apply a point-anomaly detector? Do we know if there's any effective anomaly detection algorithm that can't be equivalent to the aforementioned method, namely applying a point-anomaly detector to representation (as a single data point) of individual range of a time series?



1 Answer 1


Collective anomalies are not limited to sections of time. They can occur also when you are measuring the attribute on many sensors. A value can then be OK when seen in isolation (single datastream), but a collective anomaly when affecting many sensors.

I tried to explain it in this answer.

  • $\begingroup$ Thanks, @jonnor. In this case, can we simply treat multiple streams as a multi-variate time series, find a representation accordingly, and then feed the representation as data points to a point-anomaly detector? $\endgroup$
    – user159566
    Jan 27, 2020 at 20:54
  • $\begingroup$ Yes that is fine $\endgroup$
    – Jon Nordby
    Jan 27, 2020 at 22:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.