Quiz: The Advanced Round

Now for the harder round, covering trees, boosting, unsupervised learning, and real-world practice. These are the kinds of questions you'll hear in interviews. Once again, pause on each, and commit to an answer.

• Trees, boosting, clustering, real work
• Interview-level questions
• Commit to an answer

Question 1

Question 1. What makes the trees in a random forest different from each other? A, Different loss functions. B, Random rows and random features. C, Different programming languages. D, Different labels. Pause the video and choose your answer. The answer is B. Bootstrapped rows and random features give the diversity that makes averaging work.

Question 2

Question 2. In gradient boosting, what does each new tree try to predict? A, The original target. B, The errors of the model so far. C, Random noise. D, The feature names. Pause the video and choose your answer. The answer is B. Each tree fits the residuals, which shrinks what is left.

Question 3

Question 3. You lower the boosting learning rate. What should you usually do as well? A, Use more trees. B, Use fewer trees. C, Remove features. D, Skip validation. Pause the video and choose your answer. The answer is A. Smaller steps need more of them, and early stopping can choose how many.

Question 4

Question 4. K means finds poor groups on two crescent moon shapes. What would you try? A, More restarts only. B, DBSCAN. C, Linear regression. D, Lower the learning rate. Pause the video and choose your answer. The answer is B. DBSCAN follows density, so it can trace curved shapes.

Question 5

Question 5. What does a principal component represent? A, The most important original column. B, A direction of maximum variance. C, A cluster centre. D, A decision rule. Pause the video and choose your answer. The answer is B. Components are new directions, mixed from the original columns.

Question 6

Question 6. Which metric suits a rare event best? A, Accuracy. B, R squared. C, Precision recall AUC. D, Mean squared error. Pause the video and choose your answer. The answer is C. It focuses on the rare positive class, and ignores the easy majority.

Question 7

Question 7. You're forecasting sales. Which validation is right? A, Shuffled K fold. B, Leave one out at random. C, Walk forward by time. D, Test on the training set. Pause the video and choose your answer. The answer is C. Training must always come before testing in time.

Question 8

Question 8. What does a SHAP value tell you? A, The training time. B, How much a feature pushed one prediction. C, The best learning rate. D, The number of trees. Pause the video and choose your answer. The answer is B. SHAP divides the credit for a single prediction between features, fairly.

Question 9

Question 9. Your model's inputs look the same, but the true outcomes have changed. What's this called? A, Data drift. B, Concept drift. C, Overfitting. D, Label smoothing. Pause the video and choose your answer. The answer is B. The relationship has changed, so the same inputs now mean something different.

Question 10

Question 10. Your best tuned model scores slightly above a simple one, but it's ten times harder to run. What should you do? A, Always ship the complex one. B, Prefer the simple one unless the gain is worth it. C, Delete the data. D, Ignore both. Pause the video and choose your answer. The answer is B. Complexity has a cost, so it must earn its place.

Summary