ADM Model Learning
Hi everyone,
I have a question to understand the default behavior of the ADM (Adaptive Decision Manager) model learning process. When we send the capture response, we have two options:
Option 1: Call the Out-of-the-Box (OOTB) data flow "DF_CaptureInteraction". Option 2: Call another OOTB data flow "pxHandleResponses" based on the event strategy.
In the ADM Model, the "View" outcome is considered negative, and the "Accept" outcome is considered positive.
Now, let's consider a requirement where we have an ADM model and we receive 5 responses for the same interaction ID and subject ID in the sequence of outcomes from the channel as follows: "View, View, View, Accept, View".
In Option 1, it will write all 4 negatives and 1 positive as they are received.
In Option 2 (based on the event strategy delay learning window), it should only write the last response, which is "View" (Negative). In this example, it will have 0 positives and 1 negative.
My question is, will there be any significant impact on the model learning? Option 1 writes every capture response (both positives and negatives), while Option 2 only writes the last response within the delay learning window.
Thank you!
Pegasystems Inc.
NL
Hi @Nizam,
There will be a difference in model learning in those two scenarios. In option 2, the model will not learn properly if all (or most) interactions follow the same structure where the positive response is replaced with a negative one. If the model does not "see" the positive responses it cannot properly learn.
Usually the "pxHandleResponses" flow is used to send delayed responses in scenarios where you do not expect an explicit reject. An example is an outbound email campaign. In that scenario this feature can be used to immediately queue up a negative response, which will be emitted after a set interval (e.g. 5 days), and will be replaced by a positive response in case one is received within this window.
I hope this helps.
Best, Tom
Tech Mahindra
IN
@TomVanDuistLet me ask my question in another way. We are getting two responses from the channel, "View" and "Accept," in the same order. In option 1, it will have both positive and negative responses, while option 2 will have only positive responses with no negative responses. In this scenario, do you think option 1 works better than option 2? I was thinking that for a positive response, why should we feed a negative response again? Will it not impact the learning?