Bias is error from a model too simple to capture the real signal; variance is error from a model so flexible it fits noise in the training sample. Total expected error is bias squared plus variance plus irreducible noise. In practice you see it as a gap: high training and validation error means bias, low training error with much worse validation error means variance.
Why interviewers ask this
This separates people who memorized a definition from people who use it to debug. The interviewer wants to know what you do next after you diagnose which side you are on, because the fixes are opposite: more capacity and better features for bias, more data and regularization for variance. They are also checking that you compare training error to validation error rather than staring at one number.
How to structure your answer
- Define both terms in one sentence each.
- Explain how the training versus validation gap tells you which one dominates.
- Name the opposite fixes for each side.
- Give a concrete tuning example such as tree depth.
Example answer
Bias is the error you get because your model cannot represent the true relationship, variance is the error you get because it is chasing noise in whatever sample you happened to train on. What makes it useful is that it is a debugging tool, not a slogan. I look at training error next to validation error. If both are bad, I have a bias problem, so I add capacity, richer features, or interactions. If training error is near zero and validation is much worse, that is variance, so I pull back with stronger regularization, fewer features, or more data if I can get it. The clearest example I have hit was a gradient boosted model on transaction data. At depth ten it scored almost perfectly on training and far worse on holdout. Dropping to depth four and adding a leaf penalty closed most of the gap and the holdout AUC actually went up. The other thing I watch is that more data helps variance a lot and helps bias almost not at all, so if the learning curve has flattened, buying more rows is a waste of money.
Walking into this interview soon? GhostPilot listens to your live call, spots the question the moment it is asked, and puts a structured answer on your screen in real time. Try it on your next mock, or grab a $29 Session Pass, no subscription, for the real thing.
See how it worksFollow-up questions to expect
- How would you tell from a learning curve whether more data will help?
- What does bagging do to bias and variance, and what does boosting do?
- Can a model have high bias and high variance at the same time?
Related data scientist questions
Your interviewer will ask their own version of this. Paste your actual job description into the free Question Predictor and get the 20 questions that role is most likely to ask, with what each one is really probing.
Predict my questions