Exporting historical data

This demo shows you how to export the customer interaction data that is used by adaptive models to make predictions, including all predictor data and associated outcomes, for offline analysis.

U+ Bank has implemented Pega Customer Decision Hub™ to show a personalized banner on their website that advertises credit card offers.

When a customer is eligible for multiple credit cards, adaptive models decide which card to show. When the customer ignores the banner, the adaptive model that drives the decision regards this as negative behavior. When a customer clicks on the banner, the model regards this as a positive behavior.

As a data scientist, you may want to inspect the raw predictor data used by an adaptive model and the customer interaction outcome to validate data assumptions and check for concept drift. You can also use the data to build various predictive models externally.

All models are managed in Prediction Studio. The adaptive model that drives the decision over which banner to display is the Web Click Through Rate model.

To extract the data, you enable the recording of historical data for a selected adaptive model. A web banner typically has a low click-through rate and a significantly lower number of positive responses than negative responses. In such cases, you can sample all positive outcomes and just one percent of the negative outcomes to limit the storage space needed.

The sample percentages determine the likelihood that a customer response is recorded. The system stores the predictor data and outcome as a JSON file in a repository of your choice. By default, the data is stored for 30 days in the defaultstore repository. However, this repository points to a temporary directory, and a system architect should switch to a resilient repository to avoid data loss. Supported repository types include Microsoft Azure and Amazon S3.

For this demo, we use the default store repository and create a data set to export the data.

The data set is mapped to the file that contains the recorded historical data.

With these settings in place, the input data used for the prediction and associated outcome are stored in the configured data set when customers see an offer and click on an offer. The system architect can download the data set in DEV Studio.

Every record contains the predictor values used for the prediction, as well as the context and the decision properties, including the outcome of the interaction. All property names are automatically converted to comply with the JSON format.

To use the JSON file for further analysis, import the file into a third-party analytics tool. Keep in mind that when many customers visit the website, the file size becomes very large in a short time. To limit the storage space needed, you can lower the sample percentages.

You have reached the end of this demo. What did it show you?

• How to export the raw data that is used by adaptive models.
• What data is captured during a customer interaction.