Push feature#

This module defines various push classes used for model debugging.

The main classes are:

Push: Base push class
ExamplePush: Push based on example data
OverconfidencePush: Push for overconfidence cases
BorderlinePush: Push for borderline/underconfidence cases
ContributionPush: Push for high contribution features
PerturbationPush: Push for perturbation analysis

ExamplePush and its subclasses allow saving examples and generating one-sample tests. ContributionPush and PerturbationPush allow generating statistical tests and slicing functions.

The push classes allow converting to gRPC protobuf format via the to_grpc() method.

class giskard.push.Push#

Base push class.

push_title#: Title of the push

details#: List of details/actions for the push

tests#: List of tests to generate

pushkind#: Enum of push kind

class giskard.push.ExamplePush#

Push class based on example data.

Adds attributes:: saved_example: Example dataset row training_label: Ground truth label training_label_proba: Probability of ground truth label

Can convert to gRPC protobuf format.

class giskard.push.FeaturePush#

Push related to a specific feature.

Adds attributes:: feature: Feature name value: Feature value

Can convert to gRPC protobuf format.

class giskard.push.OverconfidencePush(training_label, training_label_proba, dataset_row, predicted_label, rate)#

Recommand actions for overconfidence cases.

Description:: Tag examples that are incorrect but that were classified with a high probability as the wrong label. This may indicate that the model is overconfident on this example. This may be due to spurious correlation or a data leak
Triggering event:: When we switch examples, the example is incorrect and the example is classified as overconfident. We quantify this as the difference between the largest probability assigned to a label and the probability assigned to the correct label (this will be 0 if the model made the correct prediction). If this is larger than a threshold (typically determined automatically depending on the number of classes), then the prediction is considered overconfident.
Call to action description:: Create one-sample test : This adds a test that checks if the model is overconfident on this example to a test suite. Get similar examples : This filters the debugging session to show examples with overconfidence only.
Requirements:: Ground truth label Classification models only

class giskard.push.BorderlinePush(training_label, training_label_proba, dataset_row, rate)#

Recommend actions for borderline/underconfidence cases.

Description:: Tag examples that are classified with very low confidence, indicating the model is unsure about the prediction. This may be due to inconsistent patterns or insufficient data.
Triggering event:: When we switch examples, the example is classified as underconfident. By default, we mark a prediction as underconfident when the second most probable prediction has a probability which is only less than 10% smaller than the predicted label
Call to action description:: Create one-sample test: This adds a test that checks if the model is underconfident on this example to a test suite. Get similar examples: This filters the debugging session to show examples with underconfidence only.
Requirements:: Classification models only

class giskard.push.ContributionPush(value=None, feature=None, feature_type=None, bounds=None, model_type=None, correct_prediction=None)#

Recommend actions for feature that have high SHAP values.

Description:: Tag features that have a high SHAP value for this example. This may indicate that the model is relying heavily on this feature to make the prediction.
Triggering event:: When we switch examples and the most contributing shapley’s value is really high compared to the rest of the features. We mark a feature as high contributing by computing SHAP values for all features. Then we calculates z-scores to find any significant outliers. If the z-score is above a threshold (typically determined automatically depending on the number of features), then the feature is considered high contributing.
Call to action description:: Save Slice : This will save the slice in the catalog and enable you to create tests more efficiently. Add Test to a test suite : This will add a test to a test suite to check if this slice performs better or worse than the rest of the dataset. Get similar examples : This will filter this debugging session to show examples from this slice only.
Requirements:: Numerical and Categorical features only

class giskard.push.PerturbationPush(value, feature, transformation_info: TransformationInfo)#

Recommend actions for feature that perturb the prediction.

Description:: Tag features that when perturbed, change the prediction. This may indicate that the model is sensitive to this feature.
Triggering event:: When we switch examples and the prediction changes when we perturb a feature. We mark a feature as sensitive by applying supported perturbations to each feature in the dataset. For numerical columns, we add/subtract values based on mean absolute deviation. For text columns, we apply predefined text transformations.
Call to action description:: Add to test suite : This will add a test to a test suite to check if a perturbation on this feature changes the prediction above a threshold.
Requirements:: Need Ground truth label Numerical and Text features only

Push tests catalog#

Test for difference in RMSE between a slice and full dataset.

Checks if the RMSE on a sliced subset of the data is significantly different from the full dataset based on a threshold and direction.

Can be used with pushes to test if problematic slices have worse RMSE.

Parameters:

model – Model to test
dataset – Full dataset
slicing_function – Function to slice dataset
threshold – Allowed RMSE difference
direction – Whether slice RMSE should increase or decrease
debug – Whether to return debug dataset

Returns:

TestResult with pass/fail, slice sizes, RMSE diff, debug dataset

Test for difference in F1 score between a slice and full dataset.

Checks if the F1 score on a sliced subset of the data is significantly different from the full dataset based on a threshold and direction.

Can be used with pushes to test if problematic slices have worse F1 score.

Parameters:

model – Model to test
dataset – Full dataset
slicing_function – Function to slice dataset
threshold – Allowed F1 score difference
direction – Whether slice F1 should increase or decrease
debug – Whether to return debug dataset

Returns:

TestResult with pass/fail, slice sizes, F1 diff, debug dataset

Test if underconfidence rate decreases for model on dataset.

Checks if the underconfidence rate for the model on the provided dataset is lower than the specified rate.

Parameters:

model – Model to test
dataset – Dataset to test on
rate – Target underconfidence rate to decrease from

Returns:

Test result with pass/fail and underconfidence rate metric

Return type:

TestResult

Test if overconfidence rate decreases for model on dataset.

Checks if the overconfidence rate for the model on the provided dataset is lower than the specified rate.

Parameters:

model – Model to test
dataset – Dataset to test on
rate – Target overconfidence rate to decrease from

Returns:

Test result with pass/fail and overconfidence rate metric

Return type:

TestResult

Test if model correctly predicts example.

Checks if the model’s prediction on the saved example matches the provided training label ground truth.

Parameters:

model – Model to test
saved_example – Example dataset
training_label – Ground truth label

Returns:

Test result with pass/fail

Return type:

TestResult

Test if model probability for ground truth label increases.

Checks if the predicted probability for the training label increases compared to the original training probability.

Parameters:

model – Model to test
saved_example – Example dataset
training_label – Ground truth label
training_label_proba – Original probability

Returns:

Test result with pass/fail and probability metric

Return type:

TestResult

giskard.push.push_test_catalog.catalog.one_sample_overconfidence_test(model: SuiteInput | BaseModel | None = None, saved_example: SuiteInput | Dataset | None = None) → GiskardTestMethod[source]#

One-sample overconfidence test for example.

Checks if the overconfidence rate for the model on the provided example is below a threshold.

Parameters:

model – Model to test
saved_example – Example dataset

Returns:

Test result with pass/fail and overconfidence rate metric

Return type:

TestResult

giskard.push.push_test_catalog.catalog.one_sample_underconfidence_test(model: SuiteInput | BaseModel | None = None, saved_example: SuiteInput | Dataset | None = None) → GiskardTestMethod[source]#

One-sample underconfidence test for example.

Checks if the underconfidence rate for the model on the provided example is below a threshold.

Parameters:

model – Model to test
saved_example – Example dataset

Returns:

Test result with pass/fail and underconfidence rate metric

Return type:

TestResult

Metamorphic test for numerical invariance with mean absolute deviation.

Checks if adding a value to a numerical feature keeps the prediction approximately the same using mean absolute deviation.

Parameters:

model – Model to test
dataset – Dataset to test
column_name – Name of numerical feature column
value_added – Value to add to column

Returns:

Test result with pass/fail

Return type:

TestResult