Adaptive Decision Manager

Welcome! In this video, we’ll dive into Pega’s Adaptive Decision Manager. This is the component of Pega Decision Management that predicts customer behavior, for example the probability that a customer will click a web banner on a self-service channel.

Adaptive Decision Manager offers two modeling techniques: Bayesian and Gradient boosting. Both techniques are online learners, which means they continuously learn from new outcomes.

The Naïve Bayes modeling technique in Adaptive Decision Manager consists of two main steps: model setup and continuous model updating.

In the model setup phase, you select the potential predictors, preferably from various sources, including customer profile and behavioral data.

You then deploy the model to production.

Each update of the model starts with the algorithm taking a batch of recent historical data and using it to establish the binning structure. This crucial step groups similar behaviors together, creating meaningful intervals that lower complexity, and serve as the foundation for the probability calculations.

For numeric predictors, such as age or income, Adaptive Decision Manager identifies intervals where customer behavior remains consistent. Customers aged 42 and aged 43 may have similar propensities to show target behavior and, after binning, reside in the same interval.

For symbolic predictors, such as product categories or customer segments, it combines categorical values that exhibit similar response patterns.

Feature selection follows preprocessing, focusing on identifying the most relevant predictors for the model. This step evaluates each predictor's individual correlation with the outcome using the area under the ROC curve, or AUC. Predictors must demonstrate significant correlation, exceeding a threshold of 0.52 AUC, to be included in the model.

Additionally, the system identifies groups of highly correlated predictors and selects the best representative from each group, supporting the independence assumption fundamental to Naïve Bayes modeling.

The core model building process then updates the probability counts that form the mathematical foundation of the model and are stored as the model's core parameters.

The final postprocessing step transforms the Naïve Bayes posterior probability to a propensity.

In this calibration step, the algorithm creates score intervals in such a way that the propensity for each next bin always increases to optimize the accuracy of the models.

The next-best-action strategy then uses the updated model to drive the delivery of highly personalized, relevant actions to each customer.

Adaptive Decision Manager incrementally updates the probabilities based on observed outcomes, allowing it to adapt to emerging patterns in customer behavior.

The second option, Gradient boosting, is a powerful machine learning technique that builds an ensemble of weak predictive models, decision trees, to create a strong overall model. It uses adaptive sliding windows to detect changes in customer behavior and determines when to grow and when to prune, eliminating the need to explicitly choose the adaptiveness.

The process begins with a single-node decision tree, and based on the data arriving from the data stream this is extended to more trees and branches within those trees. The decision tree splits the data into branches based on features that provide the most improvement of the predictive power. The best splits are determined by comparing how much information is gained by making a split, compared to not making it. The trees adapt, as branches are pruned, and other branches grow as new data is received.

The difference between the predictions and the actual values coming in, known as residuals, represents the error. Next, new decision trees are trained to predict these residual errors to correct the mistakes made by the previous trees and refine the overall prediction. A parameter called the learning rate controls how much each new tree contributes to the final prediction.

A maximum to the number of trees ensures that the Gradient boosting model does not become unnecessarily complex. The final propensity score is obtained by adding the scores of all trees, multiplied by the learning rate.

The higher predictive accuracy of a Gradient boosting model comes with a trade-off in transparency due to the high complexity of the model.

Although Gradient boosting models are harder to interpret, you can gain valuable insights from the predictor importance, which is the impact of an individual predictor on the model score.

That wraps up my presentation on Adaptive Decision Manager. Thanks for joining and until next time!