How are Stanford Machine Learning questions generated?

dotCreds builds Stanford Machine Learning practice questions from public exam objectives and DeepLearning.AI exam and documentation references. The questions are written for realistic study practice, not copied from exam dumps.

How are explanations sourced?

Each question includes an explanation and, when available, a source link back to the provider documentation or reference used to validate the answer. That keeps the practice tied to study material you can actually review.

The page tracks today's answered count and accuracy for the 10-question daily set, then saves a 7-day score history on this device so you can see your recent practice trend.

The site is the fastest way to start Stanford Machine Learning practice without installing anything. It is built for daily recall, quick weak-topic discovery, and source-backed explanations you can review immediately.

Free Stanford Machine Learning Practice Test

Stanford Machine Learning Practice Test

Start today's 10-question Stanford Machine Learning set with source-backed explanations, local progress, and a fresh rotation every morning.

10 Free Daily Questions Source-backed Explanations 150 Verified Questions

Questions updated at Aug 12, 2026, 3:38 PM CDT

Unlock the full ML Specialization bank

150 verified questions Exam Mode Practice Mode Detailed explanations Weak-area review No subscription - one-time unlock

Get the complete source-backed bank with Interview Questions, the full Study Guide, full Course Notes, detailed explanations, weak-area review, and exam-style practice.

Interview Questions Full Study Guide Full Course Notes Exam Mode Practice Mode Guided Course Detailed explanations Weak-area review No subscription

What you get with free practice

10 Free Questions Daily Fresh set every day from the live bank

Detailed Explanations Learn with clear source-backed answers

Track Your Progress Daily history and performance insights

Upgrade Anytime Unlock the full bank when you are ready

Unlock the full ML Specialization bank.

Get the full bank, Exam Mode, Practice Mode, question sets, random tests, readiness tracking, saved box scores, and review tools for this exam.

The PDF versions keep questions first and move the answer review, explanations, and distractor notes to the back of the file.

150 verified exam-style questions Every choice explained Exam Mode and Practice Mode Question sets and random tests Readiness score and trends Previous test box scores

You've answered 0/10 questions in today's set.

Locked: 140 more questions in the full bank.

Locked: exam simulation mode, practice mode, readiness tracking, and saved review history.

Checkout stays on this page, so you can keep practicing, unlock the full bank, and start Exam Mode or Practice Mode when you are ready.

Cheat Sheets

Why candidates miss this

The assumption that ensembles obviate evaluation data stems from a misunderstanding of how random forests are built and assessed. While ensembles can improve generalization, they still require a validation set to tune hyperparameters and prevent overfitting. Skipping evaluation would lead to a poorly performing model, as the ensemble’s effectiveness isn't inherent; it’s validated with data, just like individual trees. Likely wrong answer: Ensembles remove the need for evaluation data Review focus: Cross-Validation in Machine Learning

Interview question

Q: Imagine you're building a fraud detection system for an e-commerce platform. You've experimented with a single decision tree, but the performance is inconsistent. How would you improve the model's stability and robustness? Walk me through your thought process, considering the tradeoffs involved. Strong answer: The first thing I'd consider is moving to an ensemble method, specifically a random forest. A single decision tree can be quite sensitive to the specific training data it sees, leading to instability. By combining many trees, each trained on a slightly different subset of the data and features, we can reduce that sensitivity and create a more robust model. The key is that the errors of individual trees tend to cancel each other out. There's a tradeoff, of course; random forests are more computationally expensive to train and deploy than a single tree, and they can be harder to interpret. However, the improved accuracy and stability usually outweigh those costs in a fraud detection scenario where false negatives are very costly.

Understanding of ensemble methods
Tradeoff analysis (accuracy vs. complexity)
Ability to explain the benefit of reducing variance
Practical application of a machine learning concept

Caution: Simply stating 'use a random forest' without explaining *why* it's better or the associated tradeoffs. Also, suggesting that ensembles always outperform single trees without acknowledging the computational cost.

Question 1 Considering an introductory random forest design, why does an ensemble of trees typically exhibit greater stability compared to a single decision tree?

Answer choices

A. Combining many trees can reduce sensitivity to any one tree's mistakes, as the proposed design for the complete governed operational workflow.
B. Ensembles remove the need for evaluation data, for the stated security, delivery, and accountability requirements.
C. A single tree always has higher variance than an ensemble by definition in every case, for the required business outcome.
D. Tree ensembles work only on unlabeled data, for the described technical objective and its associated operational control requirements, for the affected environment.

Correct answer

Combining many trees can reduce sensitivity to any one tree's mistakes, as the proposed design for the complete governed operational workflow.

This item tests a core machine-learning concept. The selected answer, "Combining many trees can reduce sensitivity to any one tree's mistakes", is right because tree methods split data by impurity reduction, and ensembles reduce variance or sequentially correct errors. The other options point to adjacent modeling, evaluation, optimization, or unsupervised-learning ideas rather than the concept required by the stem.

Wrong-answer review

B. Ensembles remove the need for evaluation data, for the stated security, delivery, and accountability requirements.: This option handles a neighboring evaluation or preprocessing idea rather than the requested concept.
C. A single tree always has higher variance than an ensemble by definition in every case, for the required business outcome.: This option points to a separate optimization, modeling, or data-preparation step.
D. Tree ensembles work only on unlabeled data, for the described technical objective and its associated operational control requirements, for the affected environment.: This option follows another ML workflow path rather than the requirement described here.

Extra learning features

Objective/domain: Tree-Based Models

Source: CS229: Machine Learning syllabus

Question 2 For a collaborative filtering recommender system, how can a two-tower neural network represent users and items to learn embeddings that capture compatibility?

A. They are merged into a single extremely long string, and the system uses a decision tree to classify whether the string is positive or negative, under the stated technical, operational, and governance constraints.
B. The movie features are used as weights in the user network, which guarantees that the cost function remains strictly non-convex, for the described technical objective and its associated operational control requirements, within this design.
C. They are compressed using PCA, and recommendations are based on finding which user has the exact same coordinate as the movie in a 1D space, under organization-wide implementation-governance requirements.
D. They are passed through separate neural networks (user tower and movie tower) to output fixed-size embedding vectors, whose dot product represents the predicted compatibility, as described.

They are passed through separate neural networks (user tower and movie tower) to output fixed-size embedding vectors, whose dot product represents the predicted compatibility, as described.

Objective/domain: Recommender Systems

Source: Machine Learning Specialization - Deep Learning for Recommender Systems

Question 3 For a introductory boosting comparison, why do nonlinear activations matter in a neural network?

A. It guarantees the model will never overfit, for the described technical objective and its associated operational control requirements.
B. It allows the network to represent more complex patterns than stacked linear maps alone
C. It removes the need for training data, as the proposed design for the complete governed operational workflow.
D. It converts every task into clustering, under the described neural networks and deep learning criteria.

It allows the network to represent more complex patterns than stacked linear maps alone

Objective/domain: Neural Networks and Deep Learning

Source: CS229 Lecture Notes: Deep Learning

Question 4 For a clickstream recommender, what does implicit feedback utilize to quantify user engagement, represented as a numerical score reflecting the strength of the interaction?

A. Implicit feedback measures user actions like clicks, views, or purchases; it changes the target label y into a binary value (1 for interaction, 0 for no interaction) optimized via binary cross-entropy
B. Implicit feedback relies on survey questions sent via email; it changes y into a fractional score from -1.0 to +1.0 optimized via mean squared error, as the recommended implementation across the complete governed service lifecycle.
C. Implicit feedback uses computer vision to track user eye movements; it changes y into a 2D coordinate vector, for the described technical objective and its associated operational control requirements, for the required outcome.
D. Implicit feedback is when users write long textual reviews; it changes y into a categorical word index, for the described technical objective and its associated operational control requirements, within the proposed design.

Implicit feedback measures user actions like clicks, views, or purchases; it changes the target label y into a binary value (1 for interaction, 0 for no interaction) optimized via binary cross-entropy

Source: Machine Learning Specialization - Recommender Systems Detail

Question 5 In the context of anomaly detection with scarce positive examples, why is Gaussian anomaly detection often preferred over supervised classification?

A. Gaussian anomaly detection automatically generates new synthetic intrusion logs to balance the dataset, for the described technical objective and its associated operational control requirements, under the organization’s defined implementation and exception-management process.
B. Supervised classifiers are mathematically incapable of handling features that have continuous decimal values, for the described technical objective and its associated operational control requirements, for the stated implementation and support requirements.
C. Gaussian models are strictly faster to train because they do not compute any derivatives or parameter updates, for the described technical objective and its associated operational control requirements, for consideration.
D. Supervised classifiers require a much larger number of positive examples to learn the structure of the anomalous class, whereas anomaly detection only needs to model the normal class

Supervised classifiers require a much larger number of positive examples to learn the structure of the anomalous class, whereas anomaly detection only needs to model the normal class

Objective/domain: Unsupervised Learning

Source: Machine Learning Specialization - Anomaly Detection

Question 6 For a applied one-hot encoding task, what entropy value results from a binary split with probabilities 0.8 and 0.2?

A. -0.8 * log2(0.8) + 0.2 * log2(0.2)
B. -0.8 * log2(0.8) - 0.2 * log2(0.2), within the proposed design.
C. -1.0 * log2(1.0) - 0.0 * log2(0.0), for this task.
D. 0.8 * 0.2 * log2(0.8 / 0.2), for the stated tree-based models requirement.

-0.8 * log2(0.8) - 0.2 * log2(0.2), within the proposed design.

Source: Machine Learning Specialization - Decision Tree Impurity

Question 7 A logistic model outputs probabilities for an anomaly detection system, but a class decision is required. What is the function of a threshold?

A. To maximize the model's accuracy, under the documented operational and governance requirements.
B. To convert predicted probabilities into class labels, within this design.
C. To minimize the training error, for the stated supervised learning requirement.
D. To balance precision and recall, for the required outcome.

To convert predicted probabilities into class labels, within this design.

Objective/domain: Supervised Learning

Source: Machine Learning Specialization

Question 8 For a one-hot encoding task, training data distribution differs significantly from the expected deployment data. What potential issue should the team immediately assess?

A. The number of clusters in k-means, within the stated policy framework.
B. A missing PCA rotation matrix in every model, as described.
C. The impossibility of using validation data, for the required operational result and control objective.
D. Distribution shift between training and deployment data, for the required business outcome.

Distribution shift between training and deployment data, for the required business outcome.

Objective/domain: Practical ML Workflow

Question 9 During hyperparameter tuning for a neural network, a team observes high variance in the model's performance. What action should they take?

A. High bias (underfitting); resolve by adding polynomial features or decreasing lambda, under the documented operational and governance requirements.
B. High bias (underfitting); resolve by collecting more training data or training for more iterations, for the required operational result and control objective.
C. High variance (overfitting); resolve by collecting more training data, reducing the number of features, or increasing lambda
D. Data mismatch; resolve by replacing the validation set with training data, for the described technical objective and its associated operational control requirements, for consideration.

High variance (overfitting); resolve by collecting more training data, reducing the number of features, or increasing lambda

Objective/domain: Model Selection

Source: CS229 Lecture Notes: Regularization and Model Selection

Question 10 For a review-cycle bias-variance diagnosis, a logistic model outputs probabilities, but the product needs a class decision. What does a threshold do?

A. Both training error and validation error immediately flatten out at a high error rate and remain identical, for the affected environment.
B. Apply min-max scaling to compress the anomalous value into the exact center of the Gaussian distribution, under this approach.
C. It converts predicted probabilities into class labels, for the described technical objective and its associated operational control requirements, for the stated scenario.
D. Train six pairwise binary classifiers and use a voting system to determine which category wins the most duels

It converts predicted probabilities into class labels, for the described technical objective and its associated operational control requirements, for the stated scenario.

Stanford Machine Learning Practice Test

Unlock the full ML Specialization bank

Practice with explanations that teach.

What you get with free practice

Considering an introductory random forest design, why does an ensemble of trees typically exhibit greater stability compared to a single decision tree?

For a collaborative filtering recommender system, how can a two-tower neural network represent users and items to learn embeddings that capture compatibility?

For a introductory boosting comparison, why do nonlinear activations matter in a neural network?

For a clickstream recommender, what does implicit feedback utilize to quantify user engagement, represented as a numerical score reflecting the strength of the interaction?

In the context of anomaly detection with scarce positive examples, why is Gaussian anomaly detection often preferred over supervised classification?

For a applied one-hot encoding task, what entropy value results from a binary split with probabilities 0.8 and 0.2?

A logistic model outputs probabilities for an anomaly detection system, but a class decision is required. What is the function of a threshold?

For a one-hot encoding task, training data distribution differs significantly from the expected deployment data. What potential issue should the team immediately assess?

During hyperparameter tuning for a neural network, a team observes high variance in the model's performance. What action should they take?

For a review-cycle bias-variance diagnosis, a logistic model outputs probabilities, but the product needs a class decision. What does a threshold do?

You are viewing today’s free 10. Unlock 140 more questions.

Build your practice session

Set your target once.

Customize your session

Unlock the full ML Specialization bank.

7-day score keeper

Keep today’s practice moving

Answer choices

Correct answer

Wrong-answer review

Extra learning features

Why candidates miss this

Interview question

Answer choices

Correct answer

Answer choices

Correct answer

Answer choices

Correct answer

Answer choices

Correct answer

Answer choices

Correct answer

Answer choices

Correct answer

Answer choices

Correct answer

Answer choices

Correct answer

Answer choices

Correct answer

How are Stanford Machine Learning questions generated?

How are explanations sourced?

What score do I get?

Why use this site?