Improve the performance of your predictions through Machine Learning Operations (MLOps). Replace an ineffective model, scorecard, or field in a prediction with a high-accuracy model. Generate charts to compare the current and the candidate model in terms of receiver operating characteristic (ROC), gains, and lift. Deploy a candidate model in shadow mode to see how the model performs when working with real production data, but without impacting your business outcomes. If the model proves effective, you can deploy it as the active model.
When replacing a model in a prediction, you can upload a model file or connect to a model through a machine learning service. Pega Platform supports PMML, H2O MOJO, and Pega OXL model files, and can connect to models on Google AI Platform and Amazon SageMaker. You can also replace a model with a scorecard or field that contains a precalculated score.
Use the model update feature to:
- Make specific changes, for example, to address a Prediction Studio notification about a low-performing prediction.
- Update your models according to a schedule, for example, by uploading the latest churn model on a regular basis.
The model update feature allows data scientists to replace models in Prediction Studio in a non-production environment and start a deployment process to migrate the changes to production. Depending on your system, deployment involves different users and components.
Environments with Pega Customer Decision Hub
In environments that include Pega Customer Decision Hub™ and Pega 1:1 Operations Manager, approving a model update in Prediction Studio creates a change request in Pega 1:1 Operations Manager. The system automatically resolves the Change prediction request and packages the request into a revision. Team leads can then review the incoming model updates in Pega 1:1 Operations Manager. A revision manager verifies the revision in Revision Manager, and then either deploys the revision to production or withdraws it.
If you use the decision management feature of Pega Platform for other purposes, such as customer service or sales automation, components such as Pega 1:1 Operations Manager and Revision Manager are not present. In such environments, a system architect merges the branch that contains your model update to the application ruleset. A deployment manager can then deploy the updated ruleset to production.
For more information, see Updating active models in predictions.