Exporting selected model data for external analysis

This video shows you how to customize the export of adaptive model data for offline analysis in analytical tools such as Python and R. The publicly available GitHub repository Pega Data Scientist Tools helps to build meaningful plots and more with the exported data.

U+ Bank uses Customer Decision Hub to determine which credit card offer to show on their website when a customer logs in. For each offer, an adaptive model determines the likelihood that the customer will click on the web banner.

In Prediction Studio, Data Scientists continuously monitor the state of their predictions and the adaptive models that drive them.

For offline analysis of the adaptive model data, they can export the data from the Adaptive Decision Manager data mart. Two database tables contain the required monitoring information, and these tables populate two data sets.

The ADM snapshot data set contains snapshots that include the model ID, the model name, the configuration name, and model attributes such as the number of predictors, the model performance, and many others.

The ADM predictor data set contains snapshots of the binning of individual predictors. The data sets have the Model ID key in common.

Both tables can grow very large, but you typically need only the data for a selection of the models. For example, you may only be interested in the models for the application you are working on, or in just a particular channel. This demo shows you how to customize the export of the two data sets to your repository.

In the implementation phase of the project, you generate the artifacts required for the export of adaptive and predictive model snapshots, including the required data flows in Prediction Studio.

Data flows are scalable and resilient data pipelines that you can use to ingest, process, and move data to one or more destinations.

To limit the size of the exports, you may want to select only the data that you are interested in before exporting. For example, to select the data of models based on the Web Click Through Rate model configuration, adjust the two relevant data flows.

The ADM snapshot data export data flow exports the model snapshots to the repository. To configure the data flow to only export the relevant snapshots, add a filter component that only passes on snapshots on the condition that the model configuration name equals Web Click Through Rate.

The ADM predictor data set does not contain the model configuration name, but it does contain the model ID. To select the predictor binning snapshots for the selected models, create a data set that only contains the selected model IDs. Add the new data set as a second destination to the ADM snapshot data export data flow.

The ADM predictor data export data flow exports the predictor binning snapshots. To filter only the relevant snapshots, merge the predictor binning snapshots and the data set that contains the selected model IDs that you created for this purpose.

Add a Convert component to match the two classes of the source data set, which is a prerequisite for the merge operation. The data flow merges the data sets on the condition that the model IDs match, and the destination data set only contains the relevant predictor binning snapshots.

The actual export of the monitoring data is typically done in the Business Operations Environment. The exported data in the repository contains only the model snapshots of interest.

This demo has concluded. What did it show you?

• How to generate the monitoring database export artifacts.
• How to configure the auto-generated data flows to export a subset of the ADM data mart.
• How to trigger the monitoring database export.