Data exploration

What clause 7.4.2.2 asks

Before the system is built, the data is studied for places where bias could start. Sensitive and non-sensitive attributes are examined, and the links between them are tested for correlation and for cause. Finding a possible proxy is a separate task in the collection record. The purpose is to let a finding change the design, not only the test report.

Why it matters

Exploration is the bridge between the conditions and the model. A sensitive attribute can be a source of bias on its own, and a non-sensitive attribute can become one through its relation to a sensitive attribute. The clause asks for correlation and causation, which are different claims. A correlation is enough to justify a closer look; a causal claim needs more evidence. The clause also places the work before implementation, so the finding can change the design rather than only the test report.

What good looks like

Explore the data with the stakeholder groups in view. For each sensitive attribute, examine its distribution and its relation to the outcome. For each non-sensitive attribute, test its relation to the sensitive attributes and record whether the relation is correlation, plausible causation or neither. Do not use the exploration to select the metric; that comes in clause 9. Record the findings and the decisions they drove, including a decision to keep an attribute with a documented reason.

Evidence an assessor asks for

The assessor will ask what changed as a result of the exploration. An exploration that found nothing and changed nothing is possible, but it needs a record of what was tested.

• Attribute analysis The distribution and outcome relation for each sensitive attribute
• Relation analysis Correlation and causation findings for non-sensitive attributes
• Finding log Each finding, its strength and the design decision it drove
• Pre-implementation date Evidence that the exploration happened before implementation

Where it commonly fails

• The exploration is run after the model is in production
• Correlation is reported as causation
• Findings are recorded but no design decision follows

What to do before publication

Run the exploration before implementation. A finding that arrives after launch is a mitigation project, not a design choice.

Read across

• 7.4.1 Decisions about data
• 7.4.2.1 Collection and creation conditions
• 7.4.2.3 Data map
• 9.2.4.2 Examine data pre-processing
• 9.2.4.3 Testing

← 7.4.2.1 Collection and creation conditions7.4.2.3 Data map →

Full register / The dates / Free scope check