Model output distribution shifts. What are root causes and next steps?

WHAT THIS TESTS: This question tests whether you understand that a shift in model output scores is a symptom, not a diagnosis. Interviewers want to see if you can map that symptom to the three fundamental distribution shifts: covariate shift where the input distribution P(X) changes, label shift where the label prior P(Y) changes, and concept drift where the conditional relationship P(Y|X) changes. They also want to know if you understand operational causes like pipeline bugs, feature store skew, or degenerate feedback loops.

A GOOD ANSWER COVERS: A strong response starts by saying you would check input feature distributions first to detect covariate shift, because output scores cannot change unless inputs or the model change. Second, you would examine whether label priors have shifted, which would change the distribution of predicted probabilities even if the model is still calibrated. Third, you would investigate concept drift, where the world changed so the same inputs now map to different outputs. Fourth, you would check for system-level issues like a broken upstream pipeline, a feature store serving stale values, or a degenerate feedback loop where the model's own predictions are contaminating future training data. Finally, you would mention slicing the data by segment or time window to localize the shift.

COMMON WRONG ANSWERS: A red flag is treating data drift as a monolithic problem and immediately proposing to retrain without identifying the root cause. Another red flag is ignoring the possibility that the model itself is stale or that a software bug is causing the shift. Saying you would only look at aggregate accuracy metrics is also weak, because output distributions can shift while accuracy remains flat if the shift is symmetric.

LIKELY FOLLOW-UPS: An interviewer might ask how you would detect each shift in practice, such as using KL divergence or PSI for feature drift, or using a two-sample classifier to detect covariate shift. They might ask how you would handle a task with a long feedback loop where ground truth labels arrive slowly, or how you would distinguish label shift from concept drift when you only have predictions and no immediate labels. They might also ask what you would monitor if you cannot monitor labels directly.

ONE CONCRETE EXAMPLE: Imagine a grocery demand classifier that predicts whether an item will sell out. If the output probability distribution suddenly skews higher, you would first check weather and holiday features for covariate shift. If inputs look stable but the store recently started a nationwide promotion, that is label shift because the base rate of sellouts increased. If inputs and label priors are stable but the model was trained pre-pandemic and consumer behavior changed, that is concept drift. If the inventory team has been overriding predictions and those overrides are leaking back into training data, that is a degenerate feedback loop.

Source: huyenchip.com