Drift meaning

[ilyes495]

in the documentation,it says that a drift coefficient of 0.5 is very stable and a coefficient of 1 is not, so when we have two values, let's say 0.5 and 0.57 which one is the most stable? according to the documentation is 0.5, but I when I ran it on my dataset I got the one with 0.57 as the top coefficient.

Please if someone can explain me the concept and how I should interpret the results.

Thanks

[AxeldeRomblay]

Yes so 0.5 is the most stable feature you can get and 1. the worst. In your case, 0,57 > 0.5 so the one with 0.57 is less stable than the one if 0.5 but it still remains a stable feature since 0.57 << 1.

We can assume that features with:

• drift coeff < 0.6 are stable
• between 0.7 and 0.8 not stable
• drift coeff > 0.9 you have to delete it !!

[ilyes495]

thank you for your reply. so maybe you have to reverse sort the list, because I got the highest drift coefficient (0.6) at the top and the lowest (0.5) at the bottom

[AxeldeRomblay]

I have updated the code: now the drift measure is normalized between 0. and 1.

• 0. of drift (ie 0%)= the most stable feature you can get
• 1. (ie 100%) = the most unstable feature you can get