Standardizing numeric features helps a model do what?
A Use comparable scales
B Create target labels
C Remove all records
D Choose class names
Answer: A. Use comparable scales. Scaling places numeric features on comparable ranges, which can help many learning algorithms.
How does one-hot encoding represent categories?
A As a single average
B As binary indicator columns
C As ranked sentences
D As random labels
Answer: B. As binary indicator columns. One-hot encoding creates indicator columns that mark which category is present.
Missing-value imputation is used to handle what?
A Duplicate predictions
B Incorrect class names
C Missing entries
D Large feature counts
Answer: C. Missing entries. Imputation fills missing feature values using a chosen rule or statistic.
Preserving class proportions during a split is called what?
A Feature selection
B Random oversampling
C Data scaling
D Stratification
Answer: D. Stratification. A stratified split keeps class proportions similar across the resulting subsets.
Which metric focuses on the accuracy of positive predictions?
A Recall
B Precision
C Specificity
D Support
Answer: B. Precision. Precision is the fraction of predicted positives that are truly positive.
For actual positive cases, which metric measures how many were found?
A Recall
B Precision
C Accuracy
D Specificity
Answer: A. Recall. Recall measures the share of actual positive cases correctly identified.
Suppose 18 of 20 predictions are correct. What is accuracy?
A 80%
B 85%
C 90%
D 95%
Answer: C. 90%. Accuracy is correct predictions divided by all predictions: 18 ÷ 20 = 90%.
What table compares predicted classes with actual classes?
A Feature matrix
B Learning curve
C Confusion matrix
D Correlation table
Answer: C. Confusion matrix. A confusion matrix counts predictions by their actual and predicted classes.
F1 score combines which two measures?
A Recall and precision
B Accuracy and support
C Specificity and loss
D Error and coverage
Answer: A. Recall and precision. F1 is the harmonic mean of precision and recall.
On a highly imbalanced dataset, why can accuracy mislead?
A It ignores all labels
B A majority class can dominate
C It always equals recall
D It measures training speed
Answer: B. A majority class can dominate. Predicting the majority class can yield high accuracy while missing the minority class.
Mean-centering transforms a feature by subtracting what?
A Its largest value
B Its class label
C Its mean
D Its sample count
Answer: C. Its mean. Mean-centering subtracts the feature mean from each value.
During model selection, what does validation data help compare?
A File formats
B Target names
C Training records
D Model choices
Answer: D. Model choices. Validation data helps compare candidate models or hyperparameter settings during development.
In a machine-learning pipeline, steps are arranged as what?
A A random collection
B An ordered workflow
C A list of labels
D A final score
Answer: B. An ordered workflow. A pipeline applies preprocessing and estimation steps in a defined order.
Nominal categories have no natural order. Which representation suits them?
A One-hot indicators
B Ordinal ranks
C Raw row numbers
D Target averages
Answer: A. One-hot indicators. One-hot indicators represent categories without implying an order between them.
When one example can have several labels, the task is called what?
A Regression
B Clustering
C Multi-label classification
D Dimensionality reduction
Answer: C. Multi-label classification. Multi-label classification allows an example to belong to more than one class.
What name is given to an actual positive case predicted as negative?
A False positive
B True negative
C True positive
D False negative
Answer: D. False negative. A false negative occurs when a positive case is incorrectly predicted as negative.
What does an ROC curve compare across thresholds?
A Precision and recall
B True-positive and false-positive rates
C Loss and accuracy
D Labels and features
Answer: B. True-positive and false-positive rates. An ROC curve plots the true-positive rate against the false-positive rate at different thresholds.
High sensitivity to small data changes is called what?
A Recall
B Specificity
C Support
D Variance
Answer: D. Variance. High variance means a model is sensitive to changes in its training data.
Why shuffle records before a random train-test split?
A To add more labels
B To remove all outliers
C To scale every feature
D To reduce ordering bias
Answer: D. To reduce ordering bias. Shuffling can prevent an existing row order from biasing a random split.
Once fitted on training data, what can a supervised estimator do?
A Store only labels
B Generate predictions
C Replace missing records
D Draw a confusion matrix
Answer: B. Generate predictions. After fitting, a supervised predictor uses learned patterns to generate predictions.
How should preprocessing handle a category absent during fitting?
A An unseen category
B A lower accuracy
C A larger test set
D A repeated epoch
Answer: A. An unseen category. Categorical preprocessing needs a defined way to handle categories absent during fitting.
Why use stratified folds for classification cross-validation?
A To increase feature count
B To sort labels alphabetically
C To preserve class proportions
D To remove minority cases
Answer: C. To preserve class proportions. Stratified folds keep class proportions approximately consistent across folds.
In k-fold cross-validation, each fold is what?
A A model parameter
B A target column
C A training epoch
D One data partition
Answer: D. One data partition. A fold is one partition used in turn for validation while other partitions train the model.
When fitting a scaler before evaluation, which data should estimate its statistics?
A The full dataset
B Training data only
C The test set only
D Prediction outputs
Answer: B. Training data only. Fit preprocessing on training data only, then apply the learned transform to held-out data.
Why keep test data untouched until final evaluation?
A To estimate unseen-data performance
B To tune every parameter
C To choose training labels
D To increase sample size
Answer: A. To estimate unseen-data performance. A separate test set gives a final estimate without influencing model selection.