CDH Community Event: Adaptive Modeling Lessons from the Field
The predictions functionality in Pega Customer Decision Hub makes it easier than ever to drop adaptive models into Next-Best-Action decision logic. Some work is still needed, however, to help clients leverage the best of what CDH has to offer while providing confidence in the results.
In this session, Cheri Gaudet, Lead Consultant, CXForward shares lessons from the field on how to guide customers through an adaptive modeling project. You'll also learn how you can add value to your project by using OOTB reports.
Watch a replay
Note that Q&A from the session have been posted as replies below. Please continue the discussion there!
View the presentation slide deck (attached)
Pegasystems Inc.
US
@shiss This is one reason why it’s important to use response timeout windows, since the negative response is defined simply as the lack of a positive response.
Pegasystems Inc.
NL
@sundm If the channel passes an explicit negative feedback signal then the option on the Prediction to use the response timeout should not be used, instead to wait indefinitely should be used. Note: in both cases this is always based on delayed learning: the predictor values are stored as part of the decision results for the time specified in the dynamic system settings ADMShortTimeoutForDelayedLearning or ADMTimeoutForDelayedLearning depending on whether the channel uses the real-time container or not
Pegasystems Inc.
US
@shiss Were you ever asked if/how we can put weight on certain outcomes? E.g. some positive outcomes with a higher weight to model learning than some other positive outcomes?
Pegasystems Inc.
US
@shiss I would be curious to learn more about the use case for this. To my knowledge there is not a way to create a hierarchy of positive responses. The model will distinguish between positive and negative only.
Moreover, the core algorithm does not support putting extra weight on a class.
Pegasystems Inc.
US
@shiss This is one reason why it’s important to use response timeout windows, which automatically logs a negative response when a positive response hasn’t been received after a certain period of time.
Pegasystems Inc.
US
@shiss Models are effective across all channels, just keep in mind that they may take longer to learn in lower volume channels. They are set up the same way as in any other channel, except the positive and negative responses will be defined differently.
Every model requires sufficient data to be created. Without data, there will be no model. however, there are ways to overcome data volume problem. For example, it is possible to create a model at a higher level rather than a granular level to have combined data. It might be possible to use a mature model from another channel to get propensities for the low-volume channel while the low-volume channel is being trained.
Pegasystems Inc.
NL
@sundm Actually one of the strengths of Naive Bayes that it learns quickly. In addition, the default context splits make the models very quickly converge to the base rates.
Pegasystems Inc.
US
@shiss Evidence could be used when applying smoothing propensities, or when applying weights between propensities of two models based on their evidence. It could also be used when assessing the maturity level of the models.
Pegasystems Inc.
US
@shiss Are there any tips to identify negative responses in scenarios? In typical salesflow , most of the times we don't have clear negative response
Pegasystems Inc.
NL
@sundm Yes, that is why there is a default "no response" that gets sent to the models when there is no explicit (positive) response within a certain time.