Basics of Disaggregated Analysis¶

Responsible AI Tracker for disaggregated model analysis and comparison¶

To work toward fair, reliable, and safe model performance, AI practitioners need to be able to compare models and perform rigorous evaluation and testing that extend beyond single-score accuracy. Claims like “Model X is 90% accurate” on a given benchmark do not give insight as to how reliably a model may perform in the real world. Models can fail for some data cohorts or for certain conditions not reflected in the training data. For instance, a traffic sign detector may not operate well for unexpected inputs, despite the model having high accuracy overall.

Additionally, an update intended to improve overall model accuracy or performance for certain data cohorts can introduce failures for other cohorts. To help avoid this, teams need to continuously track and monitor model behavior to ensure updates are backward compatible and maintain user trust.

Envisioning and testing for what-if scenarios, like “What happens to the accuracy of the recognition model in a self-driving car when it is dark and snowing outside?” or “Does the loan approval model perform equally across ally ethnicity, gender, age, and education cohorts?” are essential for understanding how models may fail and informing decisions about what mitigations work best.

Responsible AI Tracker helps you compare models by showing performance metrics of your choosing for different data cohorts. You can do a 1:1 comparison for the same cohort across different models, tracking how different mitigation steps affect the given cohort. You can also easily create, manage, and compare cohorts of interest all within Responsible AI Tracker’s interface.