Making a predictive model adaptive?
I've taken the pega decisioning consulting courses via pega academy, however I am still confused why these are talked about separately. In the academy course they talk about how adaptive models are used when you don't have data. What about when I want my model to be adaptive and I have training data I can feed it to start with, so it isn't predicting based on nothing.
I currently have a predictive model that predicts based on the user's text input what pega category it should be associated with. How do I take it a step further and enable the stakeholder to see this decision output and determine what should be fed back into the predictive model?
Looking to make this as automatic as possible to reduce manual retrainings of the data.
***Moderator Edit-Vidyaranjan: Moved from PSC to Academy***
Pegasystems Inc.
NL
GarrettH,
Thanks for your post.
Did you notice that there are other posts that cover the same topic: https://collaborate.pega.com/question/train-adaptive-models-based-past-interaction-outcomes-uploading-customer-responses
Marco.
Updated: 29 Aug 2018 13:17 EDT
Cisco Systems Inc
US
Does this work for text analytics? I thought the adaptive model was only for propensity based decisioning. I am trying to prove myself wrong here. If I wanted to take an existing predictive MaxEnt Model, I thought I had to keep it up to date via the channels / interfaces training tab.
Per my understanding adaptive models are created in a one-vs-all approach, so each adaptive outcome (pega case category) we had in an adaptive model would be a separate model (one-vs-all) and whichever model had the highest propensity (F1 Score?) presumably based on the user id & search text.
I would rather this just be automatic though, and the predictive MaxEnt model is just updated automatically once the case is closed with the proper category it should be in. Is there a demo of something like this I could load up and look at the strategy / flow?
Cisco Systems Inc
US
In the grand scheme of things with ~500K+ cases created a qtr I doubt we could continuously append that much data to a model to help it train. What I would like to do however is look for the areas of biggest confusion in the model. ID when the model was wrong, or had very low confidence but was still right, and prioritize adding those cases back into the model to help it train faster.
Should I be trying to make a predicting model more adaptive or just remake the model as an adaptive model? The point here I am getting at is I want to reduce the change management overhead of micromanaging the model just to keep the data from becoming stale and laser focus on the feedback loop when model is very confused about a particular category.