Question

A data scientist at a lending company is preparing a training dataset to predict loan approvals. Before training, they must verify that the dataset does not disproportionately represent certain values of a sensitive 'gender' facet relative to the positive 'approved' label. They want a managed way to compute pre-training bias metrics such as Class Imbalance (CI) and Difference in Positive Proportions in Labels (DPL) and produce a report. Which approach requires the LEAST custom development effort?

Accepted Answer

Run a SageMaker Clarify pre-training bias analysis job, specifying the sensitive facet, facet values, and the positive label value.

Answer

Write a custom AWS Glue PySpark job that groups records by the gender column and manually computes proportion statistics.

Answer

Enable SageMaker Model Monitor with a data quality baseline job on the raw training data.

Answer

Use SageMaker Debugger rules configured to capture tensors during a preliminary training run.