Bias Monitor: Disparity and Group Metrics - Force Categorical

This ModelOp Center monitor computes disparity metrics (with respect to reference groups) and group metrics on protected classes, such as race or gender. Unlike the default monitor, this monitor will force all preotected classes to be treated as categorical and not continuous, even if it is a number.

Input Assets

Type	Number	Description
Baseline Data	0
Sample Data	1	A dataset corresponding to a slice of production data

Assumptions & Requirements

• Underlying BUSINESS_MODEL being monitored has a job json asset.
• BUSINESS_MODEL is a classification model.
• Protected classes under consideration are categorical features.
• Input data must contain:
  • 1 column with role=label (ground truth)
  • 1 column with role=score (model output)
  • At least 1 column with protected_class=true (protected attribute).

Execution

1. init function accepts the job json asset and validates the input schema (corresponding to the BUSINESS_MODEL being monitored).
2. metrics function instantiates the Bias Monitor class and uses the job json asset to set the protected_classes, label_column, and score_column.
3. The Aequitas Bias test and an Aequitas Group test for each protected class in the list of protected classes. A reference group for each protected class is chosen by default (first occurence).
4. Test results are appended to the list of bias tests to be returned by the model.

Monitor Output

{
    "bias":[
        <aequitas_bias_test_result>, <aequitas_group_test_result> for protected_class in protected_classes
    ]
}