topic Re: Oversampling in EMiner in SAS Data Science

Oversampling in EMiner

MagicalEmerald — Fri, 16 Dec 2016 16:16:44 GMT

Quite obvious my target event is rare, so I put a decision node to adjust the priors, Original proporition was 0.04/0.96 after adjust ment its 0.3/0.7. the decision tab on the decision node is 1.0 - 1.0 at digonal positions.

When I am scoring a test data the probabilites of the event that I am getting is around 0.3(event=1) and 0.7(event=0) I thougth EM do the adjustments for the probabilities and expected probabilites near 0.04 and 0.96. Is there anything I have to do to get the probilities adjusted, as I would get without the decision node?

Re: Oversampling in EMiner

WendyCzika — Fri, 16 Dec 2016 20:39:40 GMT

If what you are saying is the proportion in your sample was originally 0.04/0.96, then you oversampled to get 0.3/0.7, then in a Decisions node, you should have entered the original prior probabilities as the Adjusted Priors. Then if you look in the score code in the Score node, you should see lines like this doing the adjustment of the posterior probabilities:

*** Update Posterior Probabilities;
P_BAD1 = P_BAD1 * 0.04 / 0.3;
P_BAD0 = P_BAD0 * 0.96 / 0.7;
drop _sum; _sum = P_BAD1 + P_BAD0 ;