Active and candidate model type combinations
Use the Machine Learning Operations (MLOps) feature to add a candidate model to a prediction, either as a shadow of an active model or as the new active model that replaces the previous active model. When planning a model update, review the supported combinations of active and candidate model types to see what changes you can make.
You can replace the following active models in predictions:
- Adaptive models
- Predictive models
- Scorecards
- Fields
You can replace the active models with the following entities:
- Predictive models
- Scorecards
- Fields
The MLOps feature allows you to update outcome-based and supporting models in single-stage and two-stage predictions:
- Outcome-based models
- Models that are used in non-customized predictions.
- Supporting models
- Supporting models are those that are included in strategies and substrategies in customized predictions. You can use adaptive and predictive models, as well as scorecards, as supporting models.
You can update a model by using one of the following methods:
- Shadow mode
- Use this option to deploy a candidate model to the production environment as
a shadow of the active model. The system tracks the outcomes of the shadow
model, but does not use the outcomes in business decisions. You can monitor
how the shadow performs against production data, and use this insight to
decide whether to promote the shadow as the new active model.
Shadow mode is only available for predictive models. If an active model has a shadow, the Replace model option is disabled for that model. You need to either promote the shadow as the new active model or reject the shadow before you can use the Replace model option again on that model.
- Replace
- A replace operation allows you to deploy a candidate model to production as a replacement of the currently active model. A replace operation triggers the model update validation and approval process. This option is only supported for predictive models.
- Direct replace
- A direct replace operation does not involve an approval process. The change is saved to a branch that a system architect can merge to the application ruleset. A deployment manager can then deploy the application ruleset to production.
The following tables show which options are supported for each model type that you might want to replace. For example, if the active model type is adaptive, you can replace it with a predictive model, and deploy the candidate model in shadow mode or as a replacement for the adaptive model. A scorecard or field cannot replace an adaptive model.
When the active model type is ADAPTIVE
| Candidate model | Active model type | Shadow mode | Replace | Direct replace |
| Predictive | Outcome-based or supporting | ✓ | ✓ | Not available |
| Scorecard | Outcome-based or supporting | Not available | Not available | Not available |
| Field | Outcome-based | Not available | Not available | Not available |
When the active model type is PREDICTIVE
| Candidate model | Active model type | Shadow mode | Replace | Direct replace |
| Predictive | Outcome-based or supporting | ✓ | ✓ | Not available |
| Scorecard | Outcome-based or supporting | Not available | Not available | Not available |
| Field | Outcome-based | Not available | Not available | Not available |
When the active model type is SCORECARD
| Candidate model | Active model type | Shadow mode | Replace | Direct replace |
| Predictive | Outcome-based or supporting | ✓ | ✓ | Not available |
| Scorecard | Outcome-based or supporting | Not available | Not available | ✓ |
| Field | Outcome-based | Not available | Not available | ✓ |
When the active model type is FIELD
| Candidate model | Active model type | Shadow mode | Replace | Direct replace |
| Predictive | Outcome-based | ✓ | ✓ | Not available |
| Scorecard | Outcome-based | Not available | Not available | ✓ |
| Field | Outcome-based | Not available | Not available | ✓ |
Previous topic Understanding MLOps Next topic Model update statuses and notifications