All topics
Responsible ML
Shipping models people can trust: bias through proxy features, group fairness checks, and permutation importance for explainability.
Sort:
2 questions
JuniorTheoryCommonIf you drop the gender column from a credit model, is the model then free of gender bias?
JuniorTheoryCommon
If you drop the gender column from a credit model, is the model then free of gender bias?
No. Features correlated with the dropped one, such as postcode, first name or job title, act as proxies and carry the bias back in. Fairness is checked on outcomes: compare approval and error rates per group on held-out data.
Common mistakes
- ✗Assuming that dropping the sensitive column makes a model fair: correlated features such as postcode rebuild the attribute, so the bias survives.
- ✗Skipping per-group checks after removing the column: only comparing outcomes across groups shows whether the bias is actually gone.
Follow-up questions
- →Which per-group metrics would you compare for a credit approval model?
- →Why do you still need the protected attribute to audit the model?
JuniorTheoryCommonWhat does permutation importance measure, and what can it not tell you about a feature?
JuniorTheoryCommon
What does permutation importance measure, and what can it not tell you about a feature?
It is how much the model's score drops on held-out data when one feature's column is shuffled. It shows what the model relies on, not cause and effect, and correlated features share it: shuffle one, the model leans on its twin.
Common mistakes
- ✗Reading permutation importance as a causal effect: it shows how much the model depends on a feature, which may be a proxy or a leak, not what drives the target.
- ✗Trusting a low score for a correlated feature: when it is shuffled, the model falls back on its correlated twin, so both can look unimportant.
Follow-up questions
- →Why compute permutation importance on held-out data rather than the training set?
- →How would you handle two strongly correlated features when ranking importance?
