Skip to main content

         This documentation site is for previous versions. Visit our new documentation site for current releases.      

Model learning for new actions

Updated on September 15, 2022

In Pega Customer Decision Hub, every action is backed by an AI model. Additionally, every treatment within an action is also backed by an AI model. In the rest of this article actions mean both actions and treatments.

Pega Customer Decision Hub

The output of a model is a prediction known as a propensity. Propensity is the likelihood of a customer behaving favorably towards an action. The propensity output of a model could be different for different customers. A model calculates the propensity based on various parameters including but not limited to customer information and their past interaction data. This prediction output of a model represents the relevancy of an action for a customer.

The AI models are self-learning, that is, they improve the quality of their prediction automatically as and when they receive feedback in the form of customer responses. The more customer responses they receive, the more reliable their predictions are. Conversely, the propensity output of models for newly introduced actions is not reliable as they would have too few or no customer responses to learn from.

There are conditions in which it is insufficient to rely on the raw propensity returned by the AI models in order to select the right action. For example, when you turn on the system for the first time or launch a new set of actions all at once, all their propensities will be the same. So, the arbitration logic will not be able to differentiate between them. Or, if some of the actions are backed by mature models and you introduce new actions, the new actions will not get an opportunity to be presented.

Pega Customer Decision Hub complements the raw output of AI models with different mechanisms to ensure that customers receive actions that are relevant to them.

Outbound model maturity

The propensity of AI models will be less reliable in the beginning. This could result in customers receiving irrelevant actions. To help ensure that customers continue to receive relevant actions on outbound channels, Pega Customer Decision Hub provides the Outbound model maturity option in Next-Best-Action Designer.

Outbound model maturity in Next-Best-Action Designer

When this option is enabled, Pega Customer Decision Hub initially targets the outbound actions only to 2% of the population who qualify for those actions. The models start to learn as they receive customer responses. As the models start to learn, the target population for corresponding actions is gradually increased. When the models receive about 200 positive responses, 100% of the population is allowed to be presented with those actions. This approach helps in a good distribution of actions as each new action is exposed to a reduced set of audience and every new action gets a fair opportunity to learn from customer responses.

The target population for an action with an immature model is determined using the formula:

Formula to determine the immature model

Where X is the percentage of starting population considered for targeting the action.

Positives in Adaptive model

When Outbound model maturity is turned off, no capping is applied to the target population for targeting outbound actions with immature models. All (100%) qualified customers are considered, which could lead to spamming the customers with irrelevant actions.

Thompson sampling

The starting propensity of new models will be 0.5 for every customer. When multiple new actions are launched, there will be many actions with the same propensity. This could result in the same actions being selected repeatedly while other actions get suppressed.

To avoid this problem, Pega Customer Decision Hub replaces the original model propensity with the propensity selected from a range that centers around the model propensity. The width of the range depends on the number of responses the model has gathered. For new actions, the propensity is randomly chosen between 0 and 1. When the model gathers more responses, the selected propensity is closer to the model propensity. This can also be described as adding a small amount of noise to the propensity. If there is a lot of uncertainty, more noise is added. If there is little uncertainty (that is, the model has received a lot of responses), there is little noise, and the returned propensity is very close to the model propensity. So, this approach effectively only applies to immature models, namely models that have not received a lot of feedback.

Thompson sampling mechanism samples from the beta distribution around the model propensity when it adds noise to that propensity. The following graph shows the mean of the beta distribution. It demonstrates, how the calculated propensity deviates as a function of the number of positive responses (number of successes) and the number of times an action is offered to customers (total number of trials). The calculated propensity gets closer to the original propensity as the model receives more positive responses.

The Beta Distribution graph
Note: Unlike Outbound model maturity, the Thompson sampling is always applied and is applicable to both inbound and outbound actions.

AI Model control group

AI models can aid in selecting relevant actions for customers. Lift is the benefit you get by using AI models for action selection versus not using AI models. Pega Customer Decision Hub uses model control group mechanism for lift measurement. This mechanism also aids in continuous exploration to provide the model learning with unbiased samples.

You define the control group when configuring a Prediction. You can either define the control group as a percentage of all customers (the customers are then chosen randomly) or choose the customers for the control group yourself by assigning a specific field value to the customers that you want to include in the control group. For customers in the control group, Pega Customer Decision Hub applies a random propensity to the actions. Whereas, for customers in the test group, customers not in the control group, the original propensity from the model is used.

Note that, all the other conditions defined in engagement policies and arbitration still apply. Therefore, this mechanism acts only as an AI model control group rather than a global control group that tests the end-to-end next best action strategy framework.

Tip: The following mechanisms are implemented in the Predict Action Propensity prediction within the next-best-action strategy framework.

Have a question? Get answers now.

Visit the Support Center to ask questions, engage in discussions, share ideas, and help others.

Did you find this content helpful?

Want to help us improve this content?

We'd prefer it if you saw us at our best. is not optimized for Internet Explorer. For the optimal experience, please use:

Close Deprecation Notice
Contact us