Anomaly Detection buckets over time?

Hi all. I have an Anomaly Job running with hourly buckets. The values rise and fall once per day. It seems to work great in tracking that pattern, and learning about weekends.

But I have a newbie question. When ML is looking at say, the 3 p.m. bucket today, is it primarily comparing against the OTHER 3 p.m.s it saw in the past? Assuming yes, do more recent 3 p.m.s count more than say, 3 p.m. from a month ago?

Thanks!

Anomaly Detection jobs do learn the trends in your data (i.e. busier during the day than at night, busier during the weekdays than on the weekends, etc.) and those trends are "factored out" of the probabilistic model that observations in your data help form.

Example of it learning a daily cycle:

So when you say:

When ML is looking at say, the 3 p.m. bucket today, is it primarily comparing against the OTHER 3 p.m.s it saw in the past?

Basically, yes

Assuming yes, do more recent 3 p.m.s count more than say, 3 p.m. from a month ago?

Also yes. More recent observations have higher "impact" on the learning than things that happened a long time ago.

Thank you, Rich. I really appreciate it!

:grinning:

This topic was automatically closed 28 days after the last reply. New replies are no longer allowed.