Adding predictors to adaptive models

Hi! I'm Iris, and today I'll guide you through enhancing adaptive models in Pega Customer Decision Hub by adding new behavioral predictors.

We'll work with U+ Bank's credit card cross-selling implementation, where I'll show you how to incorporate real-time clickstream data to improve prediction accuracy.

As a data scientist working with adaptive models, you'll often need to enhance existing predictions with new data sources.

Today's focus is on adding Financial Services clickstream summaries that capture customer website behavior patterns, which can significantly improve your model's ability to predict customer engagement.

Let me start by showing you our prediction setup.

In Prediction Studio, we're working with the "Predict Web Propensity" prediction, which calculates the likelihood that customers will click on the credit card offers.

This prediction considers multiple factors including offer type, channel, and treatment.

The prediction configuration includes several critical settings.

The response labels define "Clicked" as our target behavior and "NoResponse" as the alternative outcome.

The response timeout is set to 30 minutes, meaning if customers do not interact with an offer within that time period, the system automatically records a NoResponse.

This timing begins when the Next-Best-Action decision is made.

Alternatively, you can start the waiting period when the Action is presented to the customer in your channel.

The underlying "Web Click Through Rate GB Customer" model uses gradient boosting as its machine learning technique at the customer level.

The model updates hourly, ensuring rapid learning from new customer interactions. With historical data recording enabled, the system captures predictor values and outcomes for every decision, creating comprehensive learning datasets.

Now let's enhance our model with rich behavioral data. The key is accessing the FSClickstream page, which contains numerous behavioral predictors that provide insights into how customers interact with the website. By selecting all these fields, we're provide the model with comprehensive clickstream data that reveals customer intent and engagement patterns.

These behavioral predictors complement traditional demographic and transactional data, resulting in a more complete picture of customer behavior.

The Financial Services clickstream summaries are particularly valuable because recent web browsing information often correlates strongly with purchase intent.

To see our enhanced model in action, I'll demonstrate the customer experience on the U+ Bank website. When customers like Troy log in, they encounter personalized credit card offers based on our adaptive model's calculations.

Troy is eligible for both the Standard Card and the Rewards Card, and the system displays the most relevant offer based on current propensity scores.

Customer interactions create valuable training data.

When Troy clicks "Learn more" within the thirty-minute window, the system records both an Impression outcome when the offer displays and a Clicked outcome when he engages.

If he ignores the offer, only the Impression outcome is recorded, followed by a NoResponse after the timeout period.

Each interaction feeds back into the model learning process.

The system captures not just the final decision, but also the behavioral context leading up to that decision through the clickstream data we added.

The Model Management landing page provides comprehensive monitoring capabilities.

In the Latest responses section, you can observe both positive and negative outcomes as they occur.

Clicked offers show both Impression and Clicked outcomes, while ignored offers display Impression followed by NoResponse outcomes.

The model details reveal how many responses have been recorded since the last update, providing insight into the learning velocity and data accumulation.

The Customer Profile Viewer offers detailed insights into individual customer journeys. For Troy Murphy, we can examine his complete interaction history, behavioral data summaries, and decision history.

The Financial Services Clickstream section confirms that the system captures comprehensive website navigation patterns.

By integrating FSClickstream predictors, we've enhanced our model's understanding of customer behavior patterns, leading to improved credit card offer targeting and higher engagement rates.

The combination of real-time behavioral data with traditional customer profile information creates a comprehensive foundation for personalized decision-making.

The key to successful predictor enhancement is starting with data that is already available in your application.

Then, develop a roadmap to gradually incorporate additional data sources, and ensure that you always comply with internal policies and regulatory requirements.

This systematic approach to predictor enhancement will improve your adaptive models' accuracy and effectiveness over time.

That's it for today. See you next time.