Summary Tests
Drift
Test for differences in the distribution of the reference dataset versus the evaluation dataset. If predictions and labels are provided, measure the performance degradation caused by shifting data as well as drift in predictions and labels themselves.
Labels and predictions are not required but they improve results.
Abnormal Inputs
Check the evaluation dataset for abnormal values commonly encountered in production. If model predictions are provided, test if the observed abnormal values cause a degradation in your model’s performance.
Labels and predictions are not required but they improve results.
Subset Performance
Test that your model performs equally well across different subsets of the evaluation dataset.
Predictions are required. Labels are required for most tests.
Bias and Fairness
Test that your model does not discriminate based on protected features. You must specify these protected features as an array value for the protected_features
key in the data_info
object.
For some tests predictions, labels, and/or model are required.
Transformations
Augment your evaluation dataset with synthetic abnormal values to proactively test your pipeline’s error-handling behavior and measure the performance degradation caused by different types of abnormal values.
Model is required. Labels are not required but they improve results.
Model Performance
Test that your model performs well on the evaluation dataset and is not degrading in performance.
Predictions are required. Labels are required for exact performance metrics, otherwise approximate performance metrics will be used.
Data Cleanliness
Test for data reliability by checking that your data is consistent and complete.
Attacks
Test the robustness of your model by measuring the maximum difference in model predictions that can be caused by small perturbations to data points.
Model is required.
Subset Performance Degradation
Test that your model’s performance on different subsets of the data has not degraded.
Predictions are required. Labels are required for most tests.
Data Poisoning Detection
Test for the presence of potentially corrupted samples in your data.
Labels are required.
Note that the default behavior of this test in Continuous Testing depends on the quality of data used for the initial stress test. If corrupted samples were detected in the stress test, then the continuous test will be more permissive.