cancel
Showing results for 
Search instead for 
Did you mean: 

Feature Effect understanding in Anomaly Detection

Feature Effect understanding in Anomaly Detection

Hello,

I'm not sure when I read 'Feature Effect' chart in Anomaly Detection, wheter I have right understanding or not.

 

In this chart, I can understand that 'mortor_2_rpm' feature should keep value from 3.15 to 3.2, because this range make the lowest anomaly score around 0.00064. Am I right? 

f_efft.PNG

0 Kudos
1 Solution

Accepted Solutions

@cookie_yamyam Feature effects is more relevant for standard supervised learning modeling as it is a partial dependence calculation.

In anomaly detection one is trying to identify outliers or rare instances and the frequency of specific values of interest are probably quite small hence the partial dependence or Feature Effects calculation does not seem applicable.

You may wish to attend the instructor-led Auto-ML classes (listed below) to get a better understanding of how the Feature Effects is derived.

https://university.datarobot.com/automl-i

 

View solution in original post

3 Replies
dalilaB
DataRobot Alumni

I did talk to my colleague, and this is what he said:

It varies in the 5th decimal place of the anomaly score..I don't think FE has great value for anomaly detection, and I would therefore not draw any strong conclusion out of it. but maybe someone can prove me wrong.

 

Ummm..

 

I read the reply "YES" from someone(It was DataRobot emplyee.) which was deleted.

After reading the reply, I delivered my interpret to our customer...

But it was wrong, right?

 

Our customer asked us what is the meaning FE dependence chart with actual and predicted lines on Anomaly Detection and Classification.

I always have trouble with interpret FE problems.

I understand this chart shows the relationship with one feature and a target.

What kind of insights I can get on Anomaly Detection and Classification problems?

0 Kudos

@cookie_yamyam Feature effects is more relevant for standard supervised learning modeling as it is a partial dependence calculation.

In anomaly detection one is trying to identify outliers or rare instances and the frequency of specific values of interest are probably quite small hence the partial dependence or Feature Effects calculation does not seem applicable.

You may wish to attend the instructor-led Auto-ML classes (listed below) to get a better understanding of how the Feature Effects is derived.

https://university.datarobot.com/automl-i