Hi, I'm trying to work with the anomaly detector API, and I'm getting some results I don't understand. On the example data, I have a very high drop but the anomaly detector fail to identify it. I've tried a lot of different combinations of sensitivity and max anomaly ratio but with no luck, the anomalies I get are on simple deviations but not on the high drop. Thanks, Alon 205885-container-mem-series1csv.txt

@Alon Mansour I think the behavior seen above is similar to this documented scenario on the batch(entire) and latest(last) point anomaly detection. An example of how this behavior detects the anomalies is shown with the same set of timeseries data in the document. Batch detection creates and applies only one model, the detection for each point is done in the context of the whole series. If the time series data trends up and down without seasonality, some points of change (dips and spikes in the data) may be missed by the model. Similarly, some points of change that are less significant than ones later in the data set may not be counted as significant enough to be incorporated into the model. To confirm I have taken the data from your file and used the dip in data to run latest(last) point API. This detection is able to find the anomaly as seen in screen shot below. Hope this helps!! If an answer is helpful, please click on or upvote which might help other community members reading this thread.

Azure Anomaly Detector API does not detect a very high drop

Accepted answer

romungi-MSFT 41,961 Reputation points Microsoft Employee

2022-05-27T10:21:30.797+00:00

@Alon Mansour I think the behavior seen above is similar to this documented scenario on the batch(entire) and latest(last) point anomaly detection. An example of how this behavior detects the anomalies is shown with the same set of timeseries data in the document.

Batch detection creates and applies only one model, the detection for each point is done in the context of the whole series. If the time series data trends up and down without seasonality, some points of change (dips and spikes in the data) may be missed by the model. Similarly, some points of change that are less significant than ones later in the data set may not be counted as significant enough to be incorporated into the model.

To confirm I have taken the data from your file and used the dip in data to run latest(last) point API. This detection is able to find the anomaly as seen in screen shot below. Hope this helps!!

If an answer is helpful, please click on or upvote which might help other community members reading this thread.
Please sign in to rate this answer.
Alon Mansour 21 Reputation points

2022-05-30T18:17:52.17+00:00

Hi @romungi-MSFT , Thank you for your answer. I've tried running the "last" API on the same data but did not get the same results as you did.
I've reduced the data to only one day (the day with the high drop) and ran the API "last" again, but it didn't find any anomalies.
Any suggestions?

Thanks,
Alon

206799-container-mem-series1-smallcsv.txt

romungi-MSFT 41,961 Reputation points Microsoft Employee

2022-05-31T04:10:52.647+00:00

@Alon Mansour I believe the data points variation is very minute with the settings for the above run. If I run the same with sensitivity of 100% the variation is detected. I have initially started the run at around 80% sensitivity and as you can see the same results are seen. When I changed the sensitivity to 100% the anomalies are detected.

For my earlier run as seen in my earlier screenshot I have used 70% sensitivity with a much larger dataset and variations. Please find the dataset for the same attached.
206907-container-mem-series2csv.txt

Alon Mansour 21 Reputation points

2022-06-02T13:31:40.237+00:00

Thank you @romungi-MSFT for your quick reply. When I ran the anomaly detection with sensitivity set to 99, it did find the anomaly I've mentioned, but since the sensitivity was so high, it also found other less significant anomalies.
What are the best practices for setting the sensitivity level?
In your replies, the first time you ran the anomaly detection on my dataset you found the anomaly with a sensitivity level of 70%, and the second time you had to increase the sensitivity level to 99%. You have mentioned that the difference was due to the number of values in the dataset (the second run was with. a smaller dataset). I don't understand this, since from what I read, Azure API does not have a state, it only uses a small window of data and calculates if the last value is anomalous. So how can the number of values change the result?
Is there an optimal number of values that needs to be sent to the API in streaming mode?

Thanks,
Alon

romungi-MSFT 41,961 Reputation points Microsoft Employee

2022-06-07T10:50:24.4+00:00

I would recommend using higher sensitivity for values with little deviation. I don't think the window of data is an issue because you are still using more than the minimum required values to be passed for the API. With more tests you can determine what sensitivity levels works best for you and try using that with your streaming data.
Sign in to comment

Azure Anomaly Detector API does not detect a very high drop

0 additional answers