データドリフトのモニタリングを設定¶

In the feature selection table:

• Features are sorted by feature importance in descending order. The secondary sort is alphabetical order by name.

• Not all features have an importance value and some features may have a negative value.

• For deployments of external models, the feature importance value is unavailable.

• For deployments of Leaderboard or custom models with training data, the table uses normalized ACE as the importance value, not Feature Impact.

• For deployments of text generation models, the table uses an assigned score—the target is 1.0, text features are 0.9, and other features are 0.1.

For all deployments with manual feature selection enabled:

• Only the selected features are tracked for drift. If a feature is not selected, it is not tracked, even if the feature exists in prediction requests or uploaded statistics.

• Switching from manual to automatic selection may result in the removal of previously tracked features from drift tracking if they are not important enough.

• Accumulated feature drift statistics are not removed by switching back and forth between automatic and manual or by adding or removing tracked features.

When manually selecting features for drift tracking:

• The available features are taken from the champion model project's feature list (if applicable) or training dataset (if applicable).

• The supported feature data types are numeric, categorical, and text.

• For deployments with imported model packages as champion, the only supported features are those with a pre-calculated baseline (in the .mlpkg file).

• For deployments with a Leaderboard model as champion:

  • For non-time series models, any supported feature in a project’s feature list is available, even if a feature is not part of model’s feature list.

  • For time series non-segmented models, any supported feature in a model’s feature list is available.

  • For time series segmented models, feature drift isn’t supported.

• For deployments with a custom model as champion:

  • For non-timeseries models, any supported column in the training data is available.

  • For time series models, any supported column in the training data is available.

• For deployments with external models, any supported column in the training data is available.

During model replacement:

• Features currently selected are removed after replacement if they are not supported or included in the replacement model.

• Feature drift is disabled if none of the currently selected features are eligible for tracking in the replacement model.

• A model replacement validation warning is displayed if either of these changes will happen after replacement.