Posted: 31 Jul 2019 4:39 EDT Last activity: 22 Aug 2019 7:15 EDT
We have some concerns while testing Adaptive model. Just to set the context we have prepared the below mentioned data and scenarios to test the Adaptive model.
1. Total Customer:36000
2. Each 12000 is having exactly same predictor data(Like age, Income etc..) and segregated in 3 category.
3. We run TrainAdaptiveModels to train the model where for the first 12000(category 1) customer we accepted a particular proposition (Lets say P1) and for rest of the category (Category 2 and 3) we rejected the same proposition(p1).
4. Now once training is done when we took one customer from 2nd segment(who Rejected P1) and found that its propensity is very high and towards the 1st set of customer which is not desirable.
5. When we check "Adaptive Model behaviour Report", only one predictor is showing as active and rest is Inactive. Performance Threshold is set .5 and all predictors are performing above 50%.
My questions are related to point 4 and point 5. Why those behaviours are happening?
Further, do we have any material on how statistically adaptive works in Pega. All the pdn meterials are functional.
***Edited by Moderator Marissa to update platform capability tags****
Although you use 36000 data points, you really only have 3 different data points. ADM will not learn anything from repeating the same data 12000 times.
So what you are observing is probably that ADM picks up on some arbitrary attribute that happens to distinguish group 2/3 from 1. Three data points is really too little data to train a model on. I usually say that ADM needs a few 100 examples with “positive behavior” (assuming that is the minority group) and with that I mean examples with variation in the data, not the same values.