ADM Model Context
Lets say I have to drive the experience for 100k offers in a store. What will happen if I create a model at group level and pass the offer name as a predictor this model?
Pegasystems Inc.
NL
@Nizam We generally do not recommend changing model context keys. It is also best practice to not add symbolic predictors with many (> 200) distinct values.
What would happen is that very likely the offer name shows up as a top predictor. What is also likely to happen is that not all offer names can be used as there is a limit of 200 unique symbols for symbolic predictors. Beyond that, they'll be grouped together. Newly introduced offers after that 200 has been reached will have no chance of ever being separated from the ones that are already lumped together, which obviously will make the model become less effective over time.
The default context configuration (issue/group/name/treatment/direction/channel) gives the models a "head start" so every model, for every treatment in a channel, will quickly figure out its base propensity then, as more data comes in, will start to individualize to customer attributes.
Hope that helps,
Tech Mahindra
IN
@Otto_PerdeckThanks for the explanation. But if I create a model at proposition(offer) level, it will create 100k models only with Issue, Group and Name Hierarchy. To avoid this, we want to group the offers and create model at group level(Business Issue, Group) and pass channel, direction as predictors. Then how do we design ADM in this scenario?
Regards,
Nizam
Pegasystems Inc.
NL
@Nizam 100 k models is a lot and would indeed call for other measures. But can I ask, what is the use case here? Retail / web shop or something?
Tech Mahindra
IN
@Otto_PerdeckAny suggestions please. Lets say we have accessories which we would like to use as propositions. The number of accessories count is more than 10k. Can we change the default ADM context to group level and limit the models?
Updated: 4 Mar 2022 7:55 EST
Pegasystems Inc.
NL
@Nizam Technically - of course you could, and if you do, you should consider adding some of the accessory / action attributes as predictors. However you would still be evaluating many accessories. In the usual 1:1 use cases the number of actions/treatments after engagement policies are applied is limited - often less than a few 100. If you are evaluating many more accessories then regardless how many models there are, this can easily become a bottleneck.
This does not look like a typical use case for 1:1 Customer Engagement. Please do make sure to engage with your Pega account executive to discuss this.
Tech Mahindra
IN
@Otto_PerdeckThanks for the comment. I am thinking like this.
1. Create all the accessories are propositions.
2. Apply Engagement policies on accessories.
3. While doing arbitration using ADM, we will not be using models created at (Issue, Group, Channel and Direction) context.
4. We will be passing attribute fields like Color, Size, Prize, Type etc. Predictors to the ADM model.
This approach works when we have propositions from different groups. If the Engagement Policy filters propositions with in the same group then point 4 might not add value to the ADM prioritization.