Interpreting AI Models for Fairness in Driverless AI
- Observability
- H2O Driverless AI
- H2O MLOps
This lesson walks through how to interpret a finished Driverless AI experiment before handing it off to end users. It covers sensitivity analysis, a what if tool, and disparate impact analysis as ways to check whether a model treats variables like gender fairly. Using a marketing model example, it shows that a fairness gap can still be an acceptable, understood limitation as long as it is documented. The lesson then moves into H2O MLOps, where you can add comments and tags to an experiment, such as noting a fairness limitation or restricting the model to marketing or demo use.
Key takeaways
- How sensitivity analysis works as a what if tool for testing a model's behavior under different variable changes
- Why a disparate impact analysis on a variable like gender can surface a fairness limitation before deployment
- The importance of documenting known model limitations so downstream devops and end users are aware of them
- How to add comments and tags in H2O MLOps to record a fairness issue and restrict a model to marketing or demo use
Start owning your AI