Quicksight Anomaly Detection Working

Q&A
, ,
1
  1. How does “Set alert past period to receive notification email with anomalies from past several time periods” work?
  2. I have setup anomaly detection to run on my SPICE dataset having enough data points on an HOURLY basis. Now the data which I get in dashboard is visists of my application every hour. So how does the schedule work , lets say every hour I can datapoints, it might be possible due to some lag in refresh 2-3 new data points come up so will the anomaly detection run for these 2-3 new data points and check if they breach threshold and then Alert me?
  3. In the above scenario if I setup Set Past Alerts as lets say so with every run will it check and give anomalies report in previous 9 runs as well because here I sometimes see that all values are not reported and the forecasted values also change
  4. Also if lets say the data which I am getting is aggregated on minute level in QUicksight and I setup anomlay runs at HOURLY (which is the minimum granularity available) then after every hour there are lot of new data points so it was giving very vague results, so does the algorithnm work for this case?
  5. With every run does Quicksight does the training every time and has stored the older datapoints and finds anomlaies for all new datapoints? Is there a limit on how many new datapoints can it evaluate in one run or does it skip some as well?
2

Hi @Saksham,

In regards to your first point, to my understanding when you select that option, you setup your anomaly detection to not only capture anomalies from within your set time period but to also capture anomalies from the number of past periods you set in this section.
So if you setup the past period alert, yes you should receive those 2-3 new data points.
Yes, it should continuously capture those previous 9 runs; AWS does not disclose how their formula for anomaly detection works though so I’m unable to provide additional information on how the forecasted values may change.
As mentioned, setting up as hourly is the smallest level of granularity currently available; if you’re getting too vague of results than it may not be ideal for your scenario.
I believe that Quick Sight uses a more continuous method of training; each run incorporates the data up to a given point and adapts instead of forgetting older points.
While I could not find any documentation that covers limits on new datapoints, I believe it’s around 1 million.

Let us know if you have any additional questions; I’ve included an informational video that explores Anomaly Detection a bit further below for your review as well.

3

I have gone thorugh all of this

I wanted to understand how does the algorithm decide what are the new points on which it needs to run anomaly detection and report them every hour, what is the logic they use behind the scene?

4

Hi @Saksham,

Here is the only additional AWS documentation that explores anomaly detection a bit further:

Ultimately though, AWS does not provide full in depth documentation on everything happening under the hood.