Answer: D. Some zero coefficients. The absolute-value penalty in L1 regularization can drive some coefficients exactly to zero.
How does L2 regularization usually affect large weights?
A Shrinks them, not usually zero
B Sets all to one
C Adds training examples
D Removes every feature
Answer: A. Shrinks them, not usually zero. L2 regularization penalizes squared weights, generally shrinking them without forcing exact zeros.
When validation performance worsens during training, what can limit overfitting?
A Add test labels
B Early stopping
C Random oversampling
D Feature hashing
Answer: B. Early stopping. Early stopping ends training when validation performance stops improving.
Why reserve validation data while tuning hyperparameters?
A To enlarge the feature set
B To label test examples
C To compare candidate settings
D To remove training records
Answer: C. To compare candidate settings. Validation results help compare settings without using the final test set for tuning.
In binary logistic regression, sigmoid output usually estimates what?
A Nearest training example
B Number of classes
C Loss gradient
D Class probability
Answer: D. Class probability. The sigmoid maps a score to an estimated probability for the positive class.
For an SVM, the margin is measured to which points?
A Training mean
B Nearest support vectors
C Class centroid
D Feature axis
Answer: B. Nearest support vectors. The margin is set by the closest training points, called support vectors.
Which ensemble averages trees trained on bootstrap samples?
A Single decision tree
B Gradient descent
C Random forest
D Linear model
Answer: C. Random forest. A random forest averages predictions from trees trained on bootstrap samples and feature subsets.
Boosting improves an ensemble mainly by doing what?
A Addressing prior learner errors
B Dropping every weak learner
C Averaging unrelated features
D Removing target labels
Answer: A. Addressing prior learner errors. Boosting trains learners sequentially so later learners address errors made earlier.
What does PCA choose as its first component direction?
A Lowest class frequency
B Largest feature label
C Shortest dataset column
D Maximum variance
Answer: D. Maximum variance. The first principal component points in the direction of greatest variance.
What does an explained-variance ratio report?
A Training examples per fold
B Fraction of total variance
C Number of selected classes
D Prediction error count
Answer: B. Fraction of total variance. The ratio states how much of the data’s total variance a component explains.
Excessive simplicity in a model commonly produces what?
A High bias
B High recall
C Data leakage
D Class imbalance
Answer: A. High bias. A model that is too simple tends to underfit and has high bias.
What does a grid search systematically test?
A Every possible model family
B Only training examples
C Specified parameter combinations
D New target labels
Answer: C. Specified parameter combinations. Grid search evaluates the parameter combinations supplied in its search grid.
For time-ordered data, which validation approach avoids future-to-past leakage?
A Random shuffle split
B Stratified random split
C Leave-one-row-out
D Chronological split
Answer: D. Chronological split. Chronological splits train on earlier data and evaluate on later data.
Why can standardizing features help gradient-based optimization?
A Features have comparable scales
B Labels become probabilities
C Every gradient becomes zero
D The test set grows
Answer: A. Features have comparable scales. Comparable feature scales can make optimization behave more efficiently.
What is a classifier’s decision boundary?
A A list of training labels
B The nearest data point
C A surface separating predicted classes
D The largest feature value
Answer: C. A surface separating predicted classes. A decision boundary separates regions assigned to different predicted classes.
Given probability 0.82 and threshold 0.50, which binary class is chosen?
A Negative class
B Positive class
C Both classes
D No class
Answer: B. Positive class. Because 0.82 exceeds the 0.50 threshold, the positive class is selected.
How does probability calibration assess a classifier?
A By counting its features
B By checking training speed
C By ranking all labels
D By comparing confidence with outcomes
Answer: D. By comparing confidence with outcomes. A calibrated model’s predicted probabilities align with observed event frequencies.
For rare positive outcomes, which curve often shows precision-recall tradeoffs clearly?
A Learning curve
B ROC curve
C Precision-recall curve
D Residual plot
Answer: C. Precision-recall curve. A precision-recall curve focuses on positive-class precision and recall, useful with rare positives.
What distinguishes feature selection from feature extraction?
A Keeps an original feature subset
B Creates only target labels
C Adds more training rows
D Changes the test answers
Answer: A. Keeps an original feature subset. Feature selection keeps chosen original features; extraction transforms inputs into new representations.
Data leakage occurs when training uses what?
A Only training examples
B Information unavailable at prediction time
C A smaller feature count
D An independent validation set
Answer: B. Information unavailable at prediction time. Leakage exposes model building to information that would not be available at prediction time.
Why should time-series splits preserve chronology?
A To randomize event order
B To balance class labels
C To prevent future-to-past leakage
D To increase feature count
Answer: C. To prevent future-to-past leakage. Preserving time order prevents later observations from influencing forecasts of earlier periods.
In five-fold cross-validation, how often is each fold held out?
A Five times
B Four times
C Two times
D Once
Answer: D. Once. Each of the five partitions serves as the held-out fold once.
Which loss is commonly used for probabilistic binary classification?
A Squared distance only
B Log loss
C Manhattan distance
D Cluster inertia
Answer: B. Log loss. Log loss evaluates predicted probabilities against binary outcomes.
Raising a binary classification threshold generally yields what?
A Fewer predicted positives
B More predicted positives
C More training records
D Fewer input features
Answer: A. Fewer predicted positives. Raising the threshold makes fewer examples qualify as positive predictions.
Feature engineering means doing what?
A Removing every feature
B Copying target labels
C Creating or transforming model inputs
D Shuffling test predictions
Answer: C. Creating or transforming model inputs. Feature engineering creates or transforms inputs to make useful patterns easier for a model to learn.