Back to Subject

Unit 1: Introduction and Data Preparation - Practice Quiz

INT234 — Predictive Analytics 60 Questions
0 Correct 0 Wrong 60 Left
0/60

1 What is the main purpose of predictive analytics?

Introduction to Predictive Analytics Easy
A. To encrypt confidential files
B. To design computer hardware
C. To store data in databases
D. To forecast likely future outcomes

2 Which type of data is commonly used to build a predictive model?

Introduction to Predictive Analytics Easy
A. Program comments
B. Random passwords
C. Historical data
D. Empty records

3 Which of the following is an example of predictive analytics?

Introduction to Predictive Analytics Easy
A. Recording a customer's current address
B. Displaying last month's total sales
C. Creating a folder for reports
D. Predicting whether a customer will leave

4 What is a predictive model?

Introduction to Predictive Analytics Easy
A. A table that stores raw records
B. A device that measures network speed
C. A system that estimates unknown outcomes
D. A chart that displays past values

5 Which field provides techniques often used in predictive analytics?

Introduction to Predictive Analytics Easy
A. Statistics
B. Cartography
C. Photography
D. Typography

6 What is machine learning?

Machine learning and its types Easy
A. Organizing files manually
B. Learning patterns from data
C. Repairing physical machines
D. Writing only fixed rules

7 Which of the following is a major type of machine learning?

Machine learning and its types Easy
A. Supervised learning
B. Sequential printing
C. Digital archiving
D. Relational formatting

8 In which type of learning does an agent learn through rewards and penalties?

Machine learning and its types Easy
A. Supervised learning
B. Reinforcement learning
C. Descriptive learning
D. Unsupervised learning

9 What does a machine learning algorithm produce after learning from training data?

Machine learning and its types Easy
A. A trained model
B. A storage device
C. A data cable
D. A network address

10 Which task predicts a continuous numerical value?

Machine learning and its types Easy
A. Association
B. Clustering
C. Classification
D. Regression

11 What kind of data is used in supervised learning?

Supervised and unsupervised learning Easy
A. Encrypted data
B. Labeled data
C. Unlabeled data
D. Duplicated data

12 Which task assigns an input to a predefined category?

Supervised and unsupervised learning Easy
A. Classification
B. Reduction
C. Clustering
D. Regression

13 What is the primary goal of unsupervised learning?

Supervised and unsupervised learning Easy
A. Predict targets from labeled data
B. Store records in sorted order
C. Replace missing values manually
D. Discover patterns in unlabeled data

14 Which technique groups similar data points together?

Supervised and unsupervised learning Easy
A. Regression
B. Clustering
C. Normalization
D. Classification

15 Predicting whether a loan applicant will default is an example of which task?

Supervised and unsupervised learning Easy
A. Clustering
B. Normalization
C. Classification
D. Aggregation

16 What is data preprocessing?

Data preprocessing Easy
A. Preparing data before model training
B. Deleting every available record
C. Presenting only final predictions
D. Designing the computer network

17 Which action is commonly used to handle missing numerical values?

Data preprocessing Easy
A. Moving them into new databases
B. Changing them into column names
C. Replacing them with the median
D. Converting them into file paths

18 Why are duplicate records often removed during data preprocessing?

Data preprocessing Easy
A. To convert numbers into text
B. To create additional variables
C. To increase the number of labels
D. To prevent repeated observations

19 What does normalization commonly do to numerical features?

Data preprocessing Easy
A. Places values on a common scale
B. Replaces values with random numbers
C. Removes all numerical columns
D. Converts values into class labels

20 Why are categorical values often encoded before model training?

Data preprocessing Easy
A. To create missing records
B. To enlarge the data files
C. To represent categories numerically
D. To remove every category

21 A telecom company wants to identify customers who are likely to cancel their subscriptions next month. Which predictive analytics task best matches this objective?

Introduction to Predictive Analytics Medium
A. Classifying customers as likely or unlikely to cancel
B. Calculating the average duration of customer calls
C. Summarizing the number of cancellations last year
D. Grouping customers by similarities in calling behavior

22 A retailer builds a model using five years of sales data, but consumer behavior changed sharply after a new competitor entered the market. What is the greatest concern when using this model?

Introduction to Predictive Analytics Medium
A. The dataset may contain too many numerical variables
B. The historical patterns may no longer represent current behavior
C. The sales records may have too many target values
D. The model may require an unsupervised learning algorithm

23 A hospital model predicts that a patient has a 0.78 probability of readmission within 30 days. How should this output be interpreted?

Introduction to Predictive Analytics Medium
A. The patient belongs to the 78th percentile of hospital stays
B. The patient will remain hospitalized for 78% of 30 days
C. The model is correct for exactly 78% of all patients
D. The model estimates a 78% chance of readmission for the patient

24 Which evaluation design best estimates how a predictive model will perform on future unseen cases?

Introduction to Predictive Analytics Medium
A. Remove records that the model predicts incorrectly
B. Evaluate it on the same records used for training
C. Evaluate it on a separate representative test set
D. Select the model with the most input variables

25 Two loan-default models have similar accuracy, but one uses information collected only after a loan is approved. Why should that model not be deployed for approval decisions?

Introduction to Predictive Analytics Medium
A. It uses an excessive number of categorical variables
B. It was evaluated with an accuracy-based performance metric
C. It contains information unavailable at prediction time
D. It predicts a target with only two possible classes

26 A delivery company wants to predict the number of minutes required for each shipment using distance, traffic, and weather data. What type of machine learning task is this?

Machine learning and its types Medium
A. Clustering, because shipments can be grouped by distance
B. Association learning, because weather and traffic are related
C. Regression, because the target is a continuous quantity
D. Classification, because delivery conditions form categories

27 An online game adjusts its strategy based on rewards received after each action. Which machine learning type is most directly represented?

Machine learning and its types Medium
A. Dimensionality reduction
B. Unsupervised learning
C. Reinforcement learning
D. Supervised learning

28 A fraud detection system is trained on transactions labeled as fraudulent or legitimate. Which statement correctly describes the learning setup?

Machine learning and its types Medium
A. It is clustering-based because transactions have similar attributes
B. It is reinforcement-based because fraud causes financial penalties
C. It is unsupervised because fraud patterns are usually uncommon
D. It is supervised because each training record has a known outcome

29 A model performs extremely well on its training data but poorly on new data. Which issue is the most likely cause?

Machine learning and its types Medium
A. Standardization
B. Overfitting
C. Underfitting
D. Label encoding

30 A company has millions of unlabeled images and a small set of labeled images. It first learns image representations from the unlabeled set and then trains a classifier with the labeled set. Which description fits this approach?

Machine learning and its types Medium
A. Rule-based classification
B. Pure supervised learning
C. Pure reinforcement learning
D. Semi-supervised learning

31 A marketing team has customer purchase histories but no predefined customer categories. It wants to discover naturally occurring customer segments. Which technique is most suitable?

Supervised and unsupervised learning Medium
A. Time-series forecasting
B. Linear regression
C. Cluster analysis
D. Binary classification

32 Which situation most clearly requires supervised learning rather than unsupervised learning?

Supervised and unsupervised learning Medium
A. Discovering groups of news articles with similar content
B. Finding products that frequently occur in the same basket
C. Predicting house prices from examples with known sale prices
D. Reducing 100 correlated variables to fewer components

33 A clustering algorithm assigns customers to four groups. What is the best way to interpret the group numbers 1, 2, 3, and 4?

Supervised and unsupervised learning Medium
A. They rank customers from least to most valuable
B. They indicate four known classes from training data
C. They are identifiers without an inherent numerical order
D. They represent predicted probabilities of membership

34 An email classifier is trained using messages labeled as spam or not spam. During deployment, a new type of spam appears that was absent from training. What is the most appropriate response?

Supervised and unsupervised learning Medium
A. Delete all earlier messages and rebuild the dataset
B. Convert the classifier directly into a clustering model
C. Replace the class labels with customer identifiers
D. Collect new labeled examples and retrain the classifier

35 A principal component analysis model converts 40 correlated measurements into 6 components while preserving most of the variation. What is its primary purpose?

Supervised and unsupervised learning Medium
A. Predicting a labeled category
B. Balancing class frequencies
C. Estimating missing targets
D. Reducing dimensionality

36 A dataset contains annual income values of 45,000, 52,000, 49,000, and one incorrectly entered value of 5,200,000. Which imputation statistic would generally be more robust if one income value were missing?

Data preprocessing Medium
A. Maximum value
B. Arithmetic mean
C. Median value
D. Value range

37 A distance-based model uses age values from 18 to 80 and annual income values from 20,000 to 500,000. Why should these features usually be scaled?

Data preprocessing Medium
A. To remove the need for a training and test split
B. To ensure that every feature becomes categorical
C. To stop the larger-scale income feature from dominating distance
D. To guarantee that all observations have unique values

38 A nominal feature named PaymentMethod has values Cash, Card, and Transfer. Which transformation is most appropriate when no natural order exists?

Data preprocessing Medium
A. Map the categories to ranks 1, 2, and 3
B. Replace each category with its character count
C. Create one binary indicator for each category
D. Normalize the category names to values from 0 to 1

39 A data scientist standardizes all records before dividing the dataset into training and test sets. What problem can this introduce?

Data preprocessing Medium
A. The training set can no longer contain numerical predictors
B. Information from the test set can influence the scaling parameters
C. Standardization can convert continuous targets into class labels
D. The test set will necessarily become larger than the training set

40 A binary classification dataset contains 98% negative cases and 2% positive cases. Which preprocessing strategy is most appropriate before training?

Data preprocessing Medium
A. Use stratified splitting and consider class rebalancing
B. Remove the target column from all training records
C. Replace every positive case with the majority label
D. Sort all records by class before creating each split

41 A hospital trains a 30-day readmission model using records from 2021–2023. Each training row includes a field indicating whether a follow-up appointment was completed within 14 days after discharge. Why is the model's validation performance likely misleading?

Introduction to Predictive Analytics Hard
A. The field creates class imbalance because most patients complete their appointments
B. The field introduces target leakage by using post-discharge information unavailable at prediction time
C. The field produces label noise because appointment completion is not the readmission outcome
D. The field causes covariate shift because appointment completion varies across calendar years

42 A fraud model is evaluated on transactions using accuracy, but only of transactions are fraudulent. A classifier that predicts every transaction as legitimate obtains accuracy. Which evaluation change best addresses the business objective of detecting fraud while controlling false alarms?

Introduction to Predictive Analytics Hard
A. Use training accuracy and select the epoch with the highest observed value
B. Use mean squared error and retain the default probability threshold of
C. Use precision-recall analysis and select a threshold under a false-alarm constraint
D. Use adjusted and increase the number of predictors until it stabilizes

43 A retailer retrains a demand model monthly. Its input distribution remains stable, but prediction error rises after a competitor begins aggressive discounting. Which diagnosis is most consistent with this observation?

Introduction to Predictive Analytics Hard
A. Label leakage changed the historical features available during model training
B. Covariate shift changed the marginal distribution of the input variables
C. Concept drift changed the relationship between existing predictors and demand
D. Sampling variance changed the model despite an unchanged data-generating process

44 Two credit-risk models have nearly identical ROC-AUC. Model X outputs well-calibrated probabilities, while Model Y systematically overestimates default risk. Loan approval depends on expected profit computed from predicted probabilities. Which statement is most accurate?

Introduction to Predictive Analytics Hard
A. The models are equivalent because ROC-AUC uniquely determines expected profitability
B. Model Y is preferable because overestimation always reduces expected credit losses
C. The models are equivalent because calibration cannot affect threshold-based decisions
D. Model X is preferable because equal ranking quality does not imply equal decision value

45 A churn model is intended to score customers every Monday. Customer behavior is strongly autocorrelated, and records from the same customer appear on many dates. Which validation design most credibly estimates future production performance?

Introduction to Predictive Analytics Hard
A. Use random row-level folds stratified by churn status across the complete history
B. Use leave-one-feature-out validation while retaining all dates in every training fold
C. Use rolling time-based splits and keep each validation period strictly after training
D. Use a single resubstitution estimate computed from the complete customer history

46 An agent learns warehouse-routing decisions by receiving delayed rewards for delivery speed and collision penalties, without being given the correct action for each state. Which learning type best describes this setup?

Machine learning and its types Hard
A. Self-supervised learning from masked states and fixed pseudo-labels alone
B. Supervised learning from state-action labels and immediate classification losses
C. Unsupervised learning from route clusters and reconstruction errors alone
D. Reinforcement learning from interaction with an environment and delayed rewards

47 A model is pretrained by predicting masked words in millions of unlabeled documents, then fine-tuned on 2,000 manually labeled complaint categories. How should the two stages be classified?

Machine learning and its types Hard
A. Semi-supervised regression followed by active learning
B. Supervised pretraining followed by unsupervised fine-tuning
C. Self-supervised pretraining followed by supervised fine-tuning
D. Unsupervised clustering followed by reinforcement learning

48 A streaming classifier is updated after each labeled transaction and gradually adapts to changing fraud patterns. Which pair of properties characterizes this training regime?

Machine learning and its types Hard
A. Online and supervised
B. Batch and reinforcement-based
C. Batch and supervised
D. Online and unsupervised

49 A company has 500 labeled failure records and 100,000 unlabeled sensor sequences from the same equipment population. Which approach most directly uses both sources during model development?

Machine learning and its types Hard
A. Semi-supervised learning using labeled and unlabeled sequences jointly
B. Unsupervised learning after permanently discarding all failure labels
C. Reinforcement learning using failure labels as environment actions
D. Supervised learning after permanently discarding all unlabeled sequences

50 A deterministic rules engine maps tax-form fields to liabilities using statutes encoded by experts. It produces accurate predictions but estimates no parameters from examples. Which statement is correct?

Machine learning and its types Hard
A. It is reinforcement learning because statutes function as delayed reward signals
B. It is predictive software but not machine learning because it does not learn from data
C. It is supervised learning because liabilities are numerical outputs derived from inputs
D. It is unsupervised learning because no labeled training dataset is required

51 Customer segments are produced by -means and then compared against known churn labels. Which statement correctly describes the training and evaluation?

Supervised and unsupervised learning Hard
A. Training is supervised because churn labels are inspected after clustering
B. Training is unsupervised, although external evaluation may use churn labels
C. Training is reinforcement-based because clusters are rewarded for label agreement
D. Training is semi-supervised because every cluster receives a numeric identifier

52 A binary classifier is trained on a case-control sample where positives were deliberately oversampled. The conditional feature distributions are representative, but the production positive rate is much lower. What is the primary consequence?

Supervised and unsupervised learning Hard
A. Both ranking and calibration remain unchanged because oversampling affects only computation time
B. Both ranking and calibration fail irreparably because class priors cannot be estimated afterward
C. The score ranking becomes invalid, but predicted probabilities remain calibrated without correction
D. The score ranking may remain useful, but predicted probabilities require prior correction or calibration

53 In a high-dimensional dataset, two analysts apply principal component analysis. Analyst A standardizes every feature; Analyst B uses raw units. One feature is measured in dollars with variance far larger than all others. What difference is most likely?

Supervised and unsupervised learning Hard
A. Analyst A's components will preserve every original pairwise Euclidean distance exactly
B. Analyst A's first component will necessarily equal the arithmetic mean of all features
C. Analyst B's first component will be dominated by the dollar-denominated feature
D. Analyst B's components will be invariant to changing dollars into thousands of dollars

54 A clustering algorithm assigns every point to a cluster even though the dataset contains compact groups separated by irregular low-density regions and includes outliers. Which method is most naturally suited to discovering such groups while labeling some points as noise?

Supervised and unsupervised learning Hard
A. Ordinary least squares with cluster identifiers as targets
B. -means with a fixed number of spherical clusters
C. DBSCAN with appropriately chosen density parameters
D. Principal component analysis using the first two loadings

55 A regression model achieves very low training error but unstable and much larger validation error across folds. Adding more predictors further lowers training error. Which intervention most directly targets the dominant problem?

Supervised and unsupervised learning Hard
A. Increase regularization or reduce model complexity
B. Evaluate only on the fold with the lowest error
C. Fit longer until the training error reaches zero
D. Decrease regularization or add interaction terms

56 A preprocessing pipeline imputes missing values with the global median and standardizes features before five-fold cross-validation. Both transformations are fitted once on the full dataset. What is the correct repair?

Data preprocessing Hard
A. Fit imputation globally but fit scaling separately within each validation fold
B. Fit imputation and scaling separately within each training fold, then transform its validation fold
C. Retain global fitting because neither transformation reads the target variable
D. Fit scaling globally but estimate imputation values from each validation fold

57 A categorical feature has 200,000 levels, many appearing once. A target-mean encoder is used with cross-validation. Which implementation best limits leakage and extreme estimates?

Data preprocessing Hard
A. Compute each row's encoding from the full dataset including that row's target
B. Assign each category its unsmoothed mean before creating cross-validation folds
C. Replace categories with arbitrary integers ordered by their observed target means
D. Compute out-of-fold encodings with smoothing toward the global target mean

58 A feature is missing more often for severely ill patients because clinicians skip a routine measurement during emergencies. The target is mortality. Which preprocessing choice preserves potentially predictive missingness without pretending the values are observed?

Data preprocessing Hard
A. Impute using training data and add an explicit missingness indicator
B. Drop all rows with missing values before constructing validation folds
C. Interpolate each missing value from records ordered by the mortality label
D. Replace missing values with the overall target mean from the full dataset

59 A dataset contains repeated measurements for each patient. Random row-level splitting places measurements from the same patient in both training and validation sets. What is the most appropriate correction?

Data preprocessing Hard
A. Shuffle rows more thoroughly before applying the same row-level split
B. Remove the patient identifier but retain overlapping patient measurements
C. Standardize measurements globally before repeating the row-level split
D. Split by patient identifier so each patient belongs to only one partition

60 A linear model uses one-hot encoding for a categorical variable with four levels and also includes an intercept. If all four indicator columns are retained, what is the key issue and standard remedy?

Data preprocessing Hard
A. Perfect multicollinearity occurs; omit one indicator or omit the intercept
B. Class imbalance occurs; oversample the least frequent indicator column
C. Target leakage occurs; estimate indicators only from validation observations
D. Heteroscedasticity occurs; normalize each indicator to unit Euclidean norm
Next MCQ
Unit 2