You ship a model and notice it performs noticeably worse for one underrepresented group than for everyone else. What could be causing that, and how would you start addressing it?
technical-conceptual · Junior level · data-ml
What the interviewer is really asking
Probes practical reasoning about diagnosing and mitigating a fairness gap — connecting the symptom to likely causes and to concrete, staged mitigation steps.
What to say
- Start by checking representation: the group may simply have too few examples, so the model never learned it well.
- Look at label quality and feature coverage for that group — labels or features may be noisier or less informative there.
- Describe staged mitigations: collect or rebalance data, then consider in-processing fairness constraints or post-processing threshold adjustments, and re-measure the gap per group.
What to avoid
- Don't dismiss it as noise without measuring the gap on a held-out, per-group slice.
- Don't jump straight to a complex fairness algorithm before checking whether the data is just thin or mislabeled for that group.
- Don't 'fix' it by silently dropping the group or the sensitive attribute and assuming the disparity disappears.
Example answers
Strong: First I'd confirm the gap is real by measuring accuracy on a held-out slice for that group, not just overall. The most common cause is under-representation — too few examples to learn from — or noisier labels and weaker features for that group. I'd start upstream by collecting or up-weighting more examples and checking the labels. If that isn't enough, I'd look at in-processing fairness constraints during training or adjusting decision thresholds per group, then re-measure the per-group gap to confirm it actually shrank.
Weak: It's probably just random variation since that group is small. I'd retrain on more total data and the numbers should even out on their own.