Examine data pre-processing
Every pre-processing step is examined for bias in the method and for bias it adds to the data set, and the effect on outcomes is recorded.
What clause 9.2.4.2 asks
The way data is prepared is studied for two things: bias in the method itself, and bias the method adds to the data. The data part of the profile supplies the context. The preparation is written down so the same steps can be repeated after retraining, and the team works out how the preparation changes what the system outputs.
Why it matters
Pre-processing is where a neutral-looking transformation can become a bias mechanism. Imputation can fill missing values differently for different groups. Scaling can compress a small group's variation. A category merge can hide a protected group inside an other bucket. The clause asks for both directions: bias in the method itself and bias the method adds. The retraining point is practical. A pipeline that is not recorded will be rebuilt differently next time, and the comparison will be invalid.
What good looks like
List every pre-processing step with its parameters and its effect on the protected attributes. Test the effect by comparing the outcome distribution before and after the step for each group. Record the version of the pipeline with the model version. When a model is retrained, re-run the same checks and compare the results with the previous run.
Evidence an assessor asks for
The assessor will ask whether the pre-processing was the same at training and at inference. A difference between the two is a silent source of bias.
- Pre-processing inventory Every step, its parameters and the fields it transforms
- Group effect analysis The outcome distribution before and after each step, by group
- Pipeline version The pipeline version tied to the model version
- Retraining comparison The pre-processing checks repeated after retraining, with the comparison
Where it commonly fails
- The pre-processing is recorded only as code
- The group effect is not tested
- Retraining rebuilds the pipeline without a comparison
What to do before publication
Record the pipeline version with the model version, and re-run the group effect check at every retraining.
Read across
- 7.4.2 Metadata
- 9.2.4 Evaluation of design and outputs: actions
- 9.2.4.3 Testing
- 9.3.4.2 Each iteration of the ongoing evaluation