Asked about The questions every ML interview asks, answered properly

Explain the bias-variance trade-off.

An answer that works

Error from a model has a part from wrong assumptions, bias, and a part from sensitivity to the particular training sample, variance. Simple models have high bias and low variance; flexible ones the reverse. The goal is the complexity that minimises total error on new data.

Then they push further

How do you tell which one you have?

Compare training and validation error. Both high and close together means bias: use a more flexible model or better features. Low training error with much higher validation error means variance: more data, regularisation or a simpler model.

Does adding more data always help?

It helps variance. It does little for bias, since a model that cannot represent the pattern will not learn it from more examples.

Reading an answer is not being able to give one. The section that teaches this is The questions every ML interview asks, answered properly, 3 hours of Supervised learning.

Learn this properly

Also asked about Supervised learning

You have a tabular dataset and a classification task. What do you try first?Why do gradient boosted trees still beat neural networks on tabular data?What is the difference between bagging and boosting?How does a decision tree choose where to split?Why does gradient boosting work?Can you trust linear regression coefficients as feature importance?Walk me through how you would approach a new tabular classification problem.Your dataset has 20% missing values in an important column. What do you do?

All 12 questions on Supervised learning