Question
Impression as negative outcome
Hi
All of our adaptive models have been setup (by our system integrator) to translate 'Impression' as a negative outcome. So let's assume a customer responds positively to an offer; Pega records a negative response (from the offer impression) as well as a positive response.
To me it does not seem correct that every time an offer is impressed it is also treated as a negative response. Firstly acceptance would be incorrect at 50% rather than 100% and secondly isn't this likely to confuse the model and yield incorrect propensities?
I can understand this works in scenarios where no positive or any other negative response is ever received, but that is not the scenario in question.
When I quizzed about this i was told 'Pega only uses the last response...' so in the example above only the positive response is considered. However I don't think this behaviour is true but at the time i was none the wiser.
Is my understanding incorrect or the implementation?
Thank you
@Otto_Perdeck thank you for this response. I have a question about the use of Impression as a negative outcome in 8.4.
In the out-of-the-box prediction, Web_Click_Through_Rate, Impression is not used as a negative outcome. However, in the Email_Click_Through_Rate prediction, Impression is one of three negative outcomes, along with Reject and NoResponse. Can you explain why Impression is included for email but not web? There is no explicit rejection of the action in either channel. (At least, in our client's case, they are not using "Reject" buttons.)
Thank you!
Cheri
Hi Cheri,
You can of course redefine the outcomes and the way they are handled. As of the why, here's a reply from one of my colleagues:
The model default Outcomes got cleaned up in 8.5 so that NoResponse is the only Negative outcome on the email model. Similar to Impressions counting as negatives on the web, the same is sometimes applied in the email channel. An open is a negative and a click on the CTA within email is recorded as the positive.
Hope that helps
Otto
For use cases where there is no explicit "negative" it is a recommended pattern to do it like you described, posting a negative for any impression then posting a "positive" when things get clicked. You are technically correct that this does, theoretically, skew the propensities a bit.
However, first of all this effect is almost always very small. Typical response rates are very low, so when propensity gets calculated as pos/(2*pos+neg) this is almost always very close to the actual number pos/(pos+neg). Secondly, we don't use the propensities as such. We (almost always) only use them in a relative way, and the ranking of the propositions is hardly affected by this "double counting". If you really, really want, there is a closed for formula to recover the unbiased propensity (divide by (1-p)) but we don't recommend doing that because it adds unnecessary complexity.
Lastly, in recent Pega versions we can deal with implicit negatives differently: you can define a "time out" on a response and only define the "positive" behavior. After the time out we set a negative if there is no positive. Since 7.4 this is a feature in CDH ("Outcome optimization") and recently, in Pega 8, this is a feature of Predictions.