Intersection Slices and Compound Groups
What you'll learn: How to evaluate model performance on overlapping subgroups to reveal hidden biases and interaction effects that single-attribute slicing misses.
The Hidden Danger of Single-Attribute Slices
When you evaluate slices like "women" or "age 65+" separately, you see part of the picture—but not the whole story. Real-world disadvantage often compounds at intersections: a model might perform acceptably on "women" overall and "elderly users" overall, yet badly underperform on "elderly women" specifically.
Intersection slices examine combinations of attributes simultaneously. Instead of just gender=female or age>=65, you analyze gender=female AND age>=65. This reveals interaction effects—where being in multiple groups simultaneously creates unique challenges your model hasn't learned to handle.
Why This Matters
Think of it like medical diagnosis: a drug might be safe for pregnant people and safe for people with diabetes, but dangerous for pregnant people with diabetes. The intersection creates a qualitatively different scenario.
In ML evaluation, you might find:
- A credit model works for high-income users and non-native speakers separately, but fails for high-income non-native speakers
- An image classifier handles dark skin tones and glasses separately, but struggles with both together
- A recommendation system serves urban users and seniors adequately alone, but poorly for urban seniors
Creating Compound Groups
Start with attributes from your Demographic and Protected Attribute Slices and Feature-Based Slicing. Create 2-way intersections first (gender × age), then 3-way if sample sizes permit (gender × age × geography). Use techniques from Error Analysis Through Slicing to identify which intersections show anomalous performance drops.
Key Takeaway: Intersection slices uncover compound disadvantages where belonging to multiple groups simultaneously creates worse outcomes than any single group membership would predict—essential for fairness and robust deployment.