Practice tests · 37 questions

Machine Learning & Data Science: the practice test

Classical ML, embeddings, metrics (precision/recall, AUC), experiment design and statistics: what Google Cloud, Databricks and Scale still test alongside GenAI. This test drills exactly that: 12 easy, 10 medium and 15 hard questions, every one explained, every explanation linking into the worked material.

Set up your test

Topic

How confident are you feeling?

Questions

12 in this pool · about 7 min

Reveal answers

Sample questions, answered

easy · sample

A bank's fraud team says missed fraud costs them 100x more than a false alarm, which analysts can clear in minutes. Which metric do you optimize first?

Accuracy, since it cleanly balances both error types in one single number

Precision, so the analysts' queue stays clean and trusted

F1, since it is the standard compromise between the two

✓Recall first, since the 100x cost asymmetry says misses dominate

Metric choice is a business decision wearing a math costume. The customer just told you the loss function: a miss costs 100x a false alarm, and false alarms have a cheap absorption path (the analyst queue). That points squarely at recall, with precision managed only enough to keep the queue workable. F1 weights both errors equally, which contradicts the stated 100x asymmetry, and accuracy is nearly meaningless at fraud's base rates. The interview skill: translate the customer's cost structure into the metric, and say the tradeoff out loud.

easy · sample

The model scored 95 percent in the pilot and 70 percent in production. Which class of cause should you rule out first?

✓Leakage or contamination: the pilot number was never real

Production hardware differences degrading numeric precision

The customer's users behaving adversarially toward the model

Normal statistical variation between any two evaluation samples

A 25-point cliff is not variance, and it is rarely hardware. The base-rate winner is that the pilot number was never real: duplicated or near-duplicated rows straddling the train-test split, a feature that quietly encodes the label (the classic being a field only filled in after the outcome), or test data used during tuning. Production is the first honest evaluation the model ever faced. The discipline this teaches: treat any surprisingly good offline number as a bug until traced, because the alternative is discovering it in front of the customer, which is exactly what happened here.

Go deeper than the quiz

A practice test measures recall. The material it draws from teaches the reasoning: