6.3 Practice Questions: Machine Learning
Key Takeaways
- Domain 2 (15-20%) centers on three decisions: model type, data split/evaluation, and which Azure ML tool to use.
- Regression predicts numbers, classification predicts known categories (supervised), and clustering discovers groups (unsupervised); classification vs. clustering hinges on whether labels are predefined.
- Training teaches, validation tunes (prevents overfitting), and test gives the final unbiased score; overfitting = high training but low test accuracy.
- Use R-squared for regression, precision/recall/F1 for classification, and silhouette score for clustering; precision avoids false positives, recall avoids false negatives.
- AutoML and Designer need no code, Notebooks need Python; real-time endpoints serve instant predictions while batch endpoints score large datasets.
Test your knowledge of Domain 2 with these practice questions covering regression, classification, clustering, deep learning, and Azure Machine Learning. After the May 2025 update, machine learning principles account for 15-20% of the exam.
What Domain 2 Tests
Most Domain 2 questions reduce to three decisions: which model type fits, how the data is split and evaluated, and which Azure ML tool to use.
Model Type (the single most-tested skill)
| The output is... | Model type | Supervised? |
|---|---|---|
| A continuous number (price, amount, temperature) | Regression | Yes |
| A known category or yes/no (spam, churn, defect) | Classification | Yes |
| Undiscovered groups in unlabeled data | Clustering | No (unsupervised) |
The trap is classification vs. clustering. The deciding question is whether the categories are predefined. If you already have labels (spam/not-spam) it is classification; if you are discovering groups with no labels, it is clustering.
Data Splits
| Dataset | Role |
|---|---|
| Training | Teaches the model the patterns |
| Validation | Tunes hyperparameters and helps prevent overfitting |
| Test | Provides the final, unbiased evaluation on unseen data |
Overfitting is high accuracy on training data but poor accuracy on test data — the model memorized rather than generalized. Underfitting is poor accuracy on both.
Evaluation Metrics
| Model type | Metrics | Direction |
|---|---|---|
| Regression | MAE, RMSE, R-squared | Errors lower; R-squared higher |
| Classification | Accuracy, Precision, Recall, F1 | Higher |
| Clustering | Silhouette score | Higher (closer to 1) |
Precision vs. recall is a classic question. Precision answers "when the model says positive, is it right?" — prioritize it to avoid false positives (e.g., do not flag legitimate email as spam). Recall answers "of all the real positives, how many did we catch?" — prioritize it to avoid false negatives (e.g., do not miss a cancer case).
Azure Machine Learning Tools
| Tool | What It Does | Code Required |
|---|---|---|
| AutoML | Automatically tries algorithms and hyperparameters | No |
| Designer | Drag-and-drop pipeline builder | No |
| Notebooks | Full code control in Python | Yes |
For deployment, a real-time endpoint serves immediate, individual predictions (low latency), while a batch endpoint scores large datasets on a schedule. "Score millions of records overnight" -> batch; "respond to each user request instantly" -> real-time.
Deep learning is tested only conceptually: it uses neural networks with multiple layers to learn complex patterns from large, often unstructured datasets (images, audio, text), and typically needs GPUs. No math or implementation is required.
Worked Reasoning: Precision vs. Recall
Precision-vs-recall trips up many candidates, so reason it out rather than memorizing. Imagine a spam filter:
- Precision = of the emails the model called spam, how many really were spam? Low precision means real emails land in the spam folder (annoying false positives).
- Recall = of all the actual spam, how much did the model catch? Low recall means spam reaches the inbox (missed false negatives).
If the cost of a false positive is high (blocking a legitimate email, or flagging a healthy patient for surgery), prioritize precision. If the cost of a false negative is high (missing fraud, or missing a cancer diagnosis), prioritize recall. The exam phrases this as a business goal — translate the goal into "which mistake is worse" and the metric follows.
Worked Reasoning: AutoML vs. Designer vs. Notebooks
Match the user's skills and goal to the tool. "No coding and no ML expertise, wants the best model automatically" -> AutoML. "Wants to visually assemble a pipeline without writing code" -> Designer. "Data scientist who needs full control in Python" -> Notebooks. If a question stresses automation and zero expertise, AutoML wins; if it stresses visual, drag-and-drop control, Designer wins.
Regression Metrics in One Line
For regression, MAE (mean absolute error) and RMSE (root mean squared error) measure error in the units of the target (lower is better; RMSE punishes large errors more heavily), and R-squared measures how much of the variance the model explains on a 0-to-1 scale (closer to 1 is better). A scenario asking "how well does the model fit overall?" points to R-squared; "how big are the typical errors, in dollars or degrees?" points to MAE or RMSE. Keep these separate from the classification metrics — mixing a regression metric into a classification scenario (or vice versa) is a common distractor.
How to Use This Practice Set
Answer the full set before reading explanations, then mark each miss by skill area. For every wrong answer, write why the correct option wins and why a tempting distractor fails — exam items retest the same idea with new wording.
Review Routine
Keep a three-column log: topic, missed rule, and the clue you should have noticed. Retake missed questions the next day, then mix them with new ones. For each ML question, first decide whether the answer is a number, a category, or a discovered group; that usually pins the model type immediately.
One more habit pays off across Domain 2: when a question describes data, immediately ask whether it is labeled or unlabeled. Labeled data with a numeric target means regression; labeled data with categories means classification; unlabeled data means clustering. Spotting the label situation first resolves the model-type question before you even read the answer choices.
A model predicts insurance claim amounts based on age, accident history, and coverage type. The predicted output is a dollar amount. This is an example of:
A data analyst wants to build a machine learning model but has no coding experience and no ML expertise. They want the system to automatically find the best algorithm for their data. Which Azure ML feature should they use?
What is the purpose of validation data in the machine learning process?
A marketing team has customer purchase data but NO predefined customer categories. They want to discover natural groups of similar customers. Which technique should they use?
Which statement about deep learning is correct?
A model has 98% accuracy on training data but only 55% accuracy on test data. What is this called?
In a spam detection model, which metric should be prioritized if the most important goal is to avoid blocking legitimate emails?
An e-commerce company needs to score 2 million customer records overnight for churn prediction. Which Azure ML deployment type should they use?