Arbitration determines how the strategy framework prioritizes the list of eligible and appropriate actions that come out of each group. The configuration and metadata for your arbitration options are stored in several decision data rules (DDR) which are used by the strategy framework.
The Next-Best-Action strategy framework includes adaptive models that learn at the action as well as the treatment level. These models apply the principles of outcome optimization at the treatment level and can be extended if you wish to apply advanced techniques or challenge the default models.
The propensity used in arbitration can be applied from the treatment level or at the action level. It is recommended to use treatment-level propensity, since it provides the best treatment on the best channel for a particular customer.
- Context weighting
Real-time contextual data is an important part of making highly relevant recommendations. Context weighting allows you to assign weighting for a specific context value to all actions within an issue or a group.
The contextual weighting definitions are stored in a decision data rule (DDR), which is managed by the Next-Best-Action Designer and used by the strategy framework.
- Business value
- You can specify the value of an action in the definition of that action. If more sophisticated calculations are needed, the extension strategy CalculateValueExtension can be customized to your specific requirements.
- Business levers
To highlight or push a particular action or group of actions, specify a weight on the individual action or add a business purpose weighting and selecting an issue only or group and then the weighting value.
These levers can be toggled on or off, which determines whether or not these weightings are applied to the final prioritization formula.
- Model maturity
For newly introduced outbound actions, AI models may not yet know which customers to target. To minimize the possibility of customers receiving actions that are not relevant, only 2% of the customer population are initially targeted with new actions. As the model collects the resulting learnings and matures, the targeted percentage increases until it reaches 100% for fully mature models. A model is considered mature after it has received 200 or more positive responses.
To prevent multiple new actions from having the same starting propensity, Pega Customer Decision Hub uses Thompson Sampling instead of the standard propensity smoothing formula. The propensity is initially set to a random value between 0 and 1. As the model matures, this quickly converges to the model propensity.