Monitor feature attribution skew and drift

This page describes how to use Vertex AI Model Monitoring with Vertex Explainable AI to detect skew and drift for the feature attributions of categorical and numerical input features.

Overview of feature attribution-based monitoring

Feature attributions indicate how much each feature in your model contributed to the predictions for each given instance. When you request predictions, you get predicted values as appropriate for your model. When you request explanations, you get the predictions along with feature attribution information.

Attribution scores are proportional to the contribution of the feature to a model's prediction. They are typically signed, indicating whether a feature helps push the prediction up or down. Attributions across all features must add up to the model's prediction score.

By monitoring feature attributions, Model Monitoring tracks changes in a feature's contributions to a model's predictions over time. A change in a key feature's attribution score often signals that the feature has changed in a way that can impact the accuracy of the model's predictions.

You can enable skew detection if you provide the original training dataset for your model; otherwise, you should enable drift detection. For more information, see Introduction to Vertex AI Model Monitoring.


To use Model Monitoring with Vertex Explainable AI, complete the following:

  1. If you are enabling skew detection, upload your training data or output of a batch explanation job for your training dataset to Cloud Storage or BigQuery. Obtain the URI link to the data. For drift detection, training data or explanation baseline is not required.

  2. Have an available model in Vertex AI that is either a tabular AutoML or imported custom training type:

Configure your model to use Vertex Explainable AI

To enable feature attribution monitoring for your custom-trained model, configure your model to use Vertex Explainable AI when you create, import, or deploy the model:

  1. Populate the ExplanationSpec.ExplanationParameters field.

  2. Set the ExplanationConfig.enableFeatureAttributes field to true.

  3. (required only for skew detection) Specify the ExplanationConfig.explanationBaseline field by providing one of the following:

Enable skew or drift detection

To set up either skew detection or drift detection, create a model deployment monitoring job:


To create a model deployment monitoring job using the Cloud console, create an endpoint:

  1. In the Google Cloud console, go to the Vertex AI Endpoints page.

    Go to Endpoints

  2. Click Create Endpoint.

  3. In the New endpoint pane, name your endpoint and set a region.

  4. Click Continue.

  5. In the Model name field, select an imported custom training or tabular AutoML model.

  6. In the Version field, select a version for your model.

  7. Click Continue.

  8. In the Model monitoring pane, make sure Enable model monitoring for this endpoint is toggled on. Any monitoring settings you configure apply to all models deployed to the endpoint.

  9. Enter a Monitoring job display name.

  10. Click Continue. The Monitoring objective pane opens, with options for skew or drift detection:

    Skew detection

    1. Select Training-serving skew detection.
    2. Under Training data source, provide a training data source.
    3. Under Target column, enter the column name from the training data that the model is trained to predict. This field is excluded from the monitoring analysis.
    4. (Optional) Under Alert thresholds, specify thresholds at which to trigger alerts. For information about how to format the thresholds, hold the pointer over the Help icon.
    5. Click Create.

    Drift detection

    1. Select Prediction drift detection.
    2. (Optional) Under Alert thresholds, specify thresholds at which to trigger alerts. For information about how to format the thresholds, hold the pointer over the Help icon.
    3. Click Create.


To create a model deployment monitoring job using the gcloud CLI, first deploy your model to an endpoint.

A monitoring job configuration applies to all deployed models under an endpoint.

Run the gcloud ai model-monitoring-jobs create command:

gcloud ai model-monitoring-jobs create \
  --project=PROJECT_ID \
  --region=REGION \
  --display-name=MONITORING_JOB_NAME \
  --endpoint=ENDPOINT_ID \
  --feature-thresholds=FEATURE_1=THRESHOLD_1,FEATURE_2=THRESHOLD_2 \
  --prediction-sampling-rate=SAMPLING_RATE \
  --monitoring-frequency=MONITORING_FREQUENCY \
  --target-field=TARGET_FIELD \


  • PROJECT_ID is the ID of your Google Cloud project. For example, my-project.

  • REGION is the location for your monitoring job. For example, us-central1.

  • MONITORING_JOB_NAME is the name of your monitoring job. For example, my-job.

  • EMAIL_ADDRESS is the email address where you want to receive alerts from Model Monitoring. For example,

  • ENDPOINT_ID is the ID of the endpoint under which your model is deployed. For example, 1234567890987654321.

  • (optional) FEATURE_1=THRESHOLD_1 is the alerting threshold for each feature you want to monitor. For example, housing-latitude=0.4. An alert is logged when the [statistical distance][stat-distance] between the input feature distribution and its corresponding baseline exceeds the specified threshold. By default, every categorical and numerical feature is monitored, with threshold values of 0.3.

  • (optional) SAMPLING_RATE is the fraction of the incoming prediction requests you want to log. For example, 0.5. If not specified, Model Monitoring logs all prediction requests.

  • (optional) MONITORING_FREQUENCY is the frequency at which you want the monitoring job to run on recently logged inputs. The minimum granularity is 1 hour. The default is 24 hours. For example, 2.

  • (required only for skew detection) TARGET_FIELD is the field that is being predicted by the model. This field is excluded from the monitoring analysis. For example, housing-price.

  • (required only for skew detection) BIGQUERY_URI is the link to the training dataset stored in BigQuery, using the following format:


    For example, bq://\my-project.\housing-data.\san-francisco.

    You can replace the bigquery-uri flag with alternative links to your training dataset:

    • For a CSV file stored in a Cloud Storage bucket, use --data-format=csv --gcs-uris=gs://BUCKET_NAME/OBJECT_NAME.

    • For a TFRecord file stored in a Cloud Storage bucket, use --data-format=tf-record --gcs-uris=gs://BUCKET_NAME/OBJECT_NAME.

    • For a [tabular AutoML managed dataset][dataset-id], use --dataset=DATASET_ID.

Python SDK

For information about the full end-to-end Model Monitoring API workflow, see the example notebook.


For more information about skew detection or drift detection, see projects.locations.modelDeploymentMonitoringJobs in the Vertex AI REST reference.

Once the monitoring job is created, Model Monitoring logs incoming prediction requests in a generated BigQuery table named PROJECT_ID.model_deployment_monitoring_ENDPOINT_ID.serving_predict. If request-response logging is enabled, Model Monitoring logs incoming requests in the same BigQuery table that is used for request-response logging.

See Using Model Monitoring for instructions on how to do the following optional tasks:

  • Update a Model Monitoring job.

  • Configure alerts for the Model Monitoring job

  • Configure alerts for anomalies.

Analyze skew and drift data

You can use the Cloud console to visualize the feature attributions of each monitored feature and learn which changes led to skew or drift. You can view the feature value distributions as a time-series or histogram.

Scorecard showing an example prediction data feature
            attribution and training data feature attribution for
            skew detection.

In a stable machine learning system, features' relative importance generally remains relatively stable over time. If an important feature drops in importance, it might signal that something about that feature has changed. Common causes of feature importance drift or skew include the following:

  • Data source changes.
  • Data schema and logging changes.
  • Changes in end user mix or behavior (for example, due to seasonal changes or outlier events).
  • Upstream changes in features generated by another machine learning model. Some examples are:
    • Model updates that cause an increase or decrease in coverage (overall or for an individual classification value).
    • A change in performance of the model (which changes the meaning of the feature).
    • Updates to the data pipeline, which can cause a decrease in overall coverage.

In addition, consider the following when analyzing feature attribution skew and drift data:

  • Track the most important features. A large change in attribution to a feature means that the feature's contribution to the prediction has changed. Because the prediction score is equal to the sum of the feature contributions, large attribution drift of the most important features usually indicates large drift in the model predictions.

  • Monitor all feature representations. Feature attributions are always numeric, regardless of the underlying feature type. Due to their additive nature, attributions to a multi-dimensional feature such as embeddings can be reduced to a single numeric value by adding up the attributions across dimensions. This allows you to use standard univariate drift detection methods for all feature types.

  • Account for feature interactions. Attribution to a feature accounts for the feature's contribution to the prediction, both individually and by its interactions with other features. If a feature's interactions with other features changes, distributions of attributions to a feature change, even if the marginal distribution of the feature remains the same.

  • Monitor feature groups. Because attributions are additive, you can add up attributions to related features to obtain the attribution of a feature group. For instance, in a credit lending model, combine the attribution to all features related to the loan type (for example, "grade", "sub_grade", "purpose") to obtain a single loan attribution. This group-level attribution can then be tracked to monitor for changes in the feature group.

What's next