Model learning for new actions
For newly introduced actions, AI models may not yet know which customers to target. To help ensure that customers receive actions that are relevant to them, Pega Customer Decision Hub uses mechanisms such as Thompson Sampling and random control group.
Thompson Sampling and model maturity
For every new outbound action, its initial probability of being accepted, which translates into the propensity value, is equal to 0.5. However, to prevent multiple new actions from having the same starting propensity, Pega Customer Decision Hub uses Thompson Sampling. This algorithm changes the propensity value to a randomly drawn number in the range between 0 and 1.
Initially, an action is presented to 2% of the population to minimize the possibility of customers receiving irrelevant offers. As the model receives feedback and matures, the targeted percentage increases until it reaches 100% for fully mature models. A model is considered mature after it has received at least 200 positive responses.
Random control group
This method always assigns a random propensity to a small percentage of the customers. This ensures that the models are exposed to responses to actions that they would not have offered without the randomization element. In this way, models keep learning and evolving. Unlike Thompson Sampling, randomization continues to happen, even if the models are already considered mature.
In addition, this mechanism provides an easy way to report on the lift of the models. In the out-of-the-box implementation, the assignment of customers to either the model or a random control group is consistent across channels but rotates every month.