Lead with the answer. Earn the depth.
A strong answer is correct, structured and responsive to context. It does not dump everything the candidate knows before answering what was asked.
Answer in 90 seconds
How would you handle class imbalance?
The eight-layer answer
Clarify context
Ask for the target, data, business cost or constraints when they change the answer.
Define directly
Give a one- or two-sentence answer before expanding.
Build intuition
Explain the idea in ordinary language.
Explain mechanics or mathematics
Show how the method produces its result.
State assumptions
Name conditions under which the method is credible.
Compare trade-offs
Discuss strengths, weaknesses and alternatives.
Apply an example
Connect the idea to a concrete dataset or decision.
Validate and monitor
Explain how you would test that the approach works in practice.
Weak versus strong
Question: Why regularize?
“Regularization prevents overfitting. L1 makes coefficients zero and L2 makes them small.”
Correct fragments, but no mechanism, assumptions, model-selection story or example.
A layered response
Regularization discourages a model from relying on excessively large coefficients. It adds a penalty to the training objective, trading a little training fit for lower variance and better generalization. L1 can create sparse coefficients, while L2 usually shrinks them smoothly. The penalty strength must be selected using validation, with preprocessing fitted inside each fold.
Handling follow-up questions
Do not defend instantly
A follow-up is often a test of flexibility, not an accusation that the first answer was wrong.
Make assumptions visible
“My answer assumes false negatives are more costly. If false positives dominate, I would change the threshold and metric.”
Say what you would measure
When uncertain, propose a discriminating experiment rather than pretending certainty.