Column drift

Definition

The column drift test allows you to select a dataset column, specify a drift detection method, and set a threshold for the drift score. Drift is measured by comparing the selected column on the current dataset with a reference dataset.

In development projects, the training set is used as the reference and the validation set as the current dataset.
In monitoring projects, the reference dataset is uploaded by the user and the production data is the current dataset.

If you want Openlayer to automatically find the best drift detection method and threshold, you can use the Feature drift, and Label drift tests instead.

Taxonomy

Task types: LLM, tabular classification, tabular regression, text classification.
Availability: and .

Why it matters

Measuring drift is crucial to maintain the relevance of your models. In development, it allows you to ensure that the data you use to validate your model is similar to the data you used to train it. In monitoring, it allows you to detect when the data your model is receiving is different from the data considered as reference.
Over time, changes in the underlying data distribution can degrade the performance of your model. Measuring drift helps in identifying these changes early, enabling timely updates or retraining of the model to maintain its performance.

Drift detection methods

One of the parameters that you must pass to the column drift test is the drift detection method. This is the method that will be used to compare the specified columns in the datasets and compute a drift score, which is what you apply a threshold to.

Openlayer supports different drift detection methods, namely:

Method	Application	Score
Anderson-Darling	Applies only to numerical columns.	Returns a p-value. If p-value < threshold, drift is detected. Recommended threshold: 0.05.
Characteristic Stability Index	Applies to categorical and numerical columns.	Returns the computed CSI value. If CSI >= threshold, drift is detected. Recommended threshold: 0.1.
Chi-Square	Applies only to categorical columns.	Returns a p-value. If p-value < threshold, drift is detected. Recommended threshold: 0.05.
Cramer-Von-Mises	Applies only to numerical columns.	Returns a p-value. If p-value < threshold, drift is detected. Recommended threshold: 0.05.
Energy Distance	Applies only to numerical columns.	Returns a distance. If distance >= threshold, drift is detected. Recommended threshold: 0.1.
Epps-Singleton	Applies only to numerical columns.	Returns a p-value. If p-value < threshold, drift is detected. Recommended threshold: 0.05.
Fisher Exact Test	Applies only to categorical columns.	Returns a p-value. If p-value < threshold, drift is detected. Recommended threshold: 0.05.
G-test	Applies only to categorical columns.	Returns a p-value. If p-value < threshold, drift is detected. Recommended threshold: 0.05.
Hellinger Distance	Applies to categorical and numerical columns.	Returns a distance. If distance >= threshold, drift is detected. Recommended threshold: 0.1.
Jensen-Shannon Distance	Applies to categorical and numerical columns.	Returns a distance. If distance >= threshold, drift is detected. Recommended threshold: 0.1.
Kullback-Leibler Divergence	Applies to categorical and numerical columns.	Returns the divergence. If divergence >= threshold, drift is detected. Recommended threshold: 0.1.
K-S Test	Applies only to numerical columns.	Returns a p-value. If p-value < threshold, drift is detected. Recommended threshold: 0.05.
Mann-Whitney U-Rank Test	Applies only to numerical columns.	Returns a p-value. If p-value < threshold, drift is detected. Recommended threshold: 0.05.
Population Stability Index	Applies to categorical and numerical columns.	Returns the computed PSI value. If PSI >= threshold, drift is detected. Recommended threshold: 0.1.
Student’s t-test	Applies only to numerical columns.	Returns a p-value. If p-value < threshold, drift is detected. Recommended threshold: 0.05.
Text Content Drift	Applies only to text columns.	Returns the ROC AUC of a binary classifier trained to distinguish text from the current and reference data. Drift is detected when the ROC AUC is high. Recommended threshold range: 0.5 - 1
Total Variation Distance	Applies only to categorical columns.	Returns a p-value. If p-value < threshold, drift is detected. Recommended threshold: 0.05.
Wasserstein Distance	Applies only to numerical columns.	Returns a distance. If distance >= threshold, drift is detected. Recommended threshold: 0.1.
Z-test	Applies only to categorical columns.	Returns a p-value. If p-value < threshold, drift is detected. Recommended threshold: 0.05.

Note that not all drift detection methods apply to all column types. For example, the “Kolmogorov-Smirnov (KS) test” is only available for numerical columns, the “Text content drift” method is only available for text columns, etc.If you select an invalid method for a column, the test will be skipped and you will see a message with the justification in the test report.String-valued columns are auto-detected as text at evaluation time, even when you list them in categoricalFeatureNames. The categorical methods in the table above (Chi-Square, Jensen-Shannon Distance, Hellinger Distance, PSI, G-test, etc.) all apply to the categorical/numerical types and will error on a text-typed column with Stattest ... isn't applicable to feature of type text. For a string column, the method that works is Text Content Drift (the only method that applies to text columns). Reserve the categorical methods for columns whose values are numeric-coded categories; use numeric methods such as K-S Test for numerical columns.

Test configuration examples

If you are writing a tests.json, here are a few valid configurations for the character length test:

[
  {
    "name": "Feature `Age` not drifted - K-S test",
    "description": "Asserts that feature `Age` has not drifted, using the K-S test with a 0.05 p-value",
    "type": "consistency",
    "subtype": "columnDrift",
    "thresholds": [
      {
        "insightName": "columnDrift",
        "insightParameters": [
          { "name": "column_name", "value": "Age" },
          { "name": "test_type", "value": "K-S Test" }
        ],
        "measurement": "driftScore",
        "operator": ">=",
        "value": 0.05
      }
    ],
    "subpopulationFilters": null,
    "mode": "development",
    "usesValidationDataset": true,
    "usesTrainingDataset": true,
    "usesMlModel": false,
    "syncId": "b4dee7dc-4f15-48ca-a282-63e2c04e0689"
  }
]

[
  {
    "name": "Feature `Age` not drifted - K-S test",
    "description": "Asserts that feature `Age` has not drifted, using the K-S test with a 0.05 p-value",
    "type": "consistency",
    "subtype": "columnDrift",
    "thresholds": [
      {
        "insightName": "columnDrift",
        "insightParameters": [
          { "name": "column_name", "value": "Age" },
          { "name": "test_type", "value": "K-S Test" }
        ],
        "measurement": "driftScore",
        "operator": ">=",
        "value": 0.05
      }
    ],
    "subpopulationFilters": null,
    "mode": "monitoring",
    "usesProductionData": true,
    "evaluationWindow": 3600,
    "delayWindow": 0,
    "syncId": "b4dee7dc-4f15-48ca-a282-63e2c04e0689"
  }
]

Get started

Workspace setup

Governance

Observability

Offline testing

Tests

Gateway

Data quality monitoring

Administration

Notifications

Other resources

Definition

Taxonomy

Why it matters

Drift detection methods

Test configuration examples

​Definition

​Taxonomy

​Why it matters

​Drift detection methods

​Test configuration examples

​Related

Definition

Taxonomy

Why it matters

Drift detection methods

Test configuration examples

Related