I85PREP
⌘ K
TASC Data Science and Digital Solutions study materials
TASC Data Science and Digital Solutions study resources

TASC Data Science and Digital Solutions practice questions

One question at a time. Build your TASC Data Science and Digital Solutions understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Data Science and Digital Solutions

In a relational database, which key uniquely identifies each row of a table and cannot contain null values?

  1. Foreign key
  2. Primary key Correct answer.
  3. Surrogate key
  4. Candidate key

Explanation

A primary key is the chosen identifier for a table; every row must have a unique, non-null value in that column or column combination. A foreign key instead references another table, a candidate key is any minimal unique key before one is chosen, and a surrogate key is a system-generated identifier that only becomes primary if selected.

Question 2Data Science and Digital Solutions

A dataset column records customer ages but contains several blank cells. What is the most appropriate term for these missing entries?

  1. Noise
  2. Missing values Correct answer.
  3. Outliers
  4. Duplicates

Explanation

Blank or absent entries in a data column are called missing values, and handling them (dropping, imputing, or flagging) is a standard cleaning step. Outliers are extreme but present observations, duplicates are repeated records, and noise is random error in recorded values rather than absence.

Question 3Data Science and Digital Solutions

Which measure of central tendency is most resistant to extreme outliers in a numeric variable?

  1. Median Correct answer.
  2. Range
  3. Mean
  4. Standard deviation

Explanation

The median depends only on the middle position of ordered values, so a few extreme observations barely move it. The mean averages every value and is pulled toward outliers, while range and standard deviation are measures of spread rather than central tendency.

Question 4Data Science and Digital Solutions

In supervised machine learning, what distinguishes classification from regression?

  1. Classification predicts a categorical label; regression predicts a numeric quantity Correct answer.
  2. Classification uses labeled data; regression uses unlabeled data
  3. Classification requires deep neural networks; regression uses linear models
  4. Classification evaluates with accuracy; regression cannot be evaluated

Explanation

The defining difference is the target type: classification assigns inputs to discrete classes, whereas regression estimates a continuous numeric output. Both are supervised tasks using labeled training data, and both can be evaluated with suitable metrics.

Question 5Data Science and Digital Solutions

Which practice best protects sensitive personal data stored in an analytics database?

  1. Storing credentials inside the analysis notebook
  2. Encrypting the data at rest and restricting access by role Correct answer.
  3. Publishing the raw data in an open dashboard
  4. Copying the data to a shared drive for convenience

Explanation

Encryption at rest plus role-based access control limits exposure if storage is compromised and enforces least privilege. The other options all widen access or leak secrets, which directly increases the risk of unauthorized disclosure.

Question 6Data Science and Digital Solutions

What does a data dictionary primarily document?

  1. The hardware specifications of the database server
  2. The visual color palette of dashboards
  3. The names, types, and meanings of fields in a dataset Correct answer.
  4. The project's staffing plan and budget

Explanation

A data dictionary is metadata describing each field: its name, data type, allowed values, units, and business meaning. This helps analysts interpret and use data consistently, whereas server specs, budgets, and design palettes are unrelated artifacts.

Question 7Data Science and Digital Solutions

Which chart type is best suited to show how a single numeric variable is distributed across many observations?

  1. Histogram Correct answer.
  2. Pie chart
  3. Scatter plot
  4. Stacked bar chart

Explanation

A histogram bins a numeric variable and displays the count or frequency per bin, revealing shape, spread, and skew. Pie charts show parts of a whole for categories, scatter plots show relationships between two variables, and stacked bars compare category compositions.

Question 8Data Science and Digital Solutions

What is the main purpose of splitting a dataset into training and test sets?

  1. To increase the number of features available to the model
  2. To reduce the total amount of data that must be stored
  3. To guarantee the model achieves perfect accuracy
  4. To estimate how well the model generalizes to unseen data Correct answer.

Explanation

Holding out data that the model never sees during training provides an unbiased estimate of performance on new observations. The split does not reduce storage needs, add features, or guarantee any particular accuracy level.

Question 9Data Science and Digital Solutions

A model flags 90 of 100 truly fraudulent transactions but also flags 500 legitimate transactions as fraud. Which metric best captures the burden placed on investigators by the false flags?

  1. AUC
  2. Precision Correct answer.
  3. Recall
  4. Sensitivity

Explanation

Precision is the share of flagged cases that are truly fraudulent, so here it is 90 divided by 590, which is low and reflects the heavy false-positive load. Recall and sensitivity measure how many true frauds were caught, and AUC summarizes ranking across thresholds rather than the practical false-alarm volume.

Question 10Data Science and Digital Solutions

An analyst computes the correlation between hours studied and exam score as 0.82. Which conclusion is justified?

  1. There is a strong positive linear association between the two variables Correct answer.
  2. Studying more hours causes higher scores
  3. About 82 percent of students passed the exam
  4. The score increases by exactly 0.82 points per study hour

Explanation

A correlation of 0.82 indicates a strong positive linear relationship, but correlation alone does not establish causation or the size of the per-hour effect. The slope would be needed to state points per hour, and correlation says nothing about pass rates.

Question 11Data Science and Digital Solutions

A table of orders stores customer city in every order row, so updating a customer's city requires editing many rows and risks inconsistency. Which design change addresses this problem?

  1. Move city into a separate customer table referenced by customer ID Correct answer.
  2. Add a second city column to record the previous value
  3. Index the city column for faster lookups
  4. Increase the length of the city column

Explanation

This is a classic update anomaly caused by storing a fact in multiple places; normalizing city into a customer table keyed by customer ID removes the duplication. Adding columns, enlarging the field, or indexing do not eliminate the repeated value.

Question 12Data Science and Digital Solutions

A survey question offers the response options 'Satisfied' and 'Very satisfied' but omits any negative option. Which survey design flaw does this illustrate?

  1. Small sample size
  2. Double-barreled wording
  3. Leading question order
  4. Unbalanced response scale Correct answer.

Explanation

An unbalanced scale offers choices skewed toward one side, forcing respondents to pick among only positive alternatives. Double-barreled items ask about two things at once, leading order refers to earlier questions biasing later ones, and sample size is unrelated to the option set.

Question 13Data Science and Digital Solutions

A team wants to detect unusual spikes in daily website traffic without any labeled history of past incidents. Which approach fits best?

  1. A/B testing two homepage designs
  2. Supervised classification trained on incident labels
  3. Unsupervised anomaly detection on the traffic series Correct answer.
  4. Linear regression predicting page views from user age

Explanation

With no labeled examples of past anomalies, the task is unsupervised: model the normal traffic pattern and flag deviations from it. Supervised classification needs labels, regression on an unrelated predictor does not detect spikes, and A/B testing compares designs rather than monitoring anomalies.

Question 14Data Science and Digital Solutions

A dataset records income in a currency where a few executives earn over 100 times the typical salary. Which summary is most informative for describing typical income?

  1. Maximum and minimum
  2. Sum and count
  3. Median and interquartile range Correct answer.
  4. Mean and standard deviation

Explanation

With severe skew from extreme earners, the median and interquartile range describe the typical value and spread without being dominated by the tail. Mean and standard deviation are pulled upward by the outliers, and max, min, sum, and count do not convey the center.

Question 15Data Science and Digital Solutions

In a star schema, what is the role of the fact table?

  1. It holds the schema's documentation
  2. It stores descriptive attributes used for filtering
  3. It stores measurable events linked to dimension keys Correct answer.
  4. It enforces referential integrity across dimensions

Explanation

A fact table records quantitative events such as sales amounts, with foreign keys pointing to dimension tables that hold descriptive attributes. Descriptive filters live in dimensions, and referential integrity is a constraint rather than the fact table's purpose.

Question 16Data Science and Digital Solutions

A pipeline reads a CSV where the 'date' column sometimes appears as 2023-05-01 and sometimes as 05/01/2023. What should the pipeline do before loading the data?

  1. Store the column as free text without parsing
  2. Drop all rows where dates appear in the second format
  3. Average the two formats into one value
  4. Standardize the date format to a single convention Correct answer.

Explanation

Mixed date formats must be parsed and converted to one consistent representation, ideally an ISO date type, so downstream comparisons and joins work correctly. Dropping valid rows loses data, storing free text prevents date operations, and averaging dates is meaningless.

Question 17Data Science and Digital Solutions

Why is stratified sampling often preferred over simple random sampling when a dataset contains a small minority class?

  1. It guarantees each stratum is sampled proportionally, preserving rare group representation Correct answer.
  2. It increases the number of features in the model
  3. It removes the need for any train-test split
  4. It eliminates all sampling bias from the study

Explanation

Stratified sampling draws from each subgroup in proportion to its size, so a rare class is reliably represented in both training and test data. It does not remove the need for splitting, cannot eliminate every source of bias, and does not add features.

Question 18Data Science and Digital Solutions

A dashboard shows monthly revenue as a single huge number with no comparison. Which improvement most helps executives judge whether performance is good?

  1. Change the number to a different font
  2. Add a comparison to target and to the same month last year Correct answer.
  3. Add more decimal places to the figure
  4. Remove the revenue figure entirely

Explanation

A bare number lacks context; comparing it to a target and to the prior year lets viewers judge whether the result is strong or weak. Font, precision, and removal do not supply the missing benchmark.

Question 19Data Science and Digital Solutions

A model scores 0.98 accuracy on a test set where 98 percent of records belong to the negative class. What does this result most likely indicate?

  1. The accuracy is misleading because the classes are imbalanced Correct answer.
  2. The model performs excellently on both classes
  3. The model has overfitted the training data
  4. The test set is too small to compute accuracy

Explanation

Predicting the majority class every time yields 98 percent accuracy on such a dataset, so the score reflects class imbalance rather than genuine skill. Metrics such as precision, recall, or balanced accuracy are needed to reveal performance on the minority class.

Question 20Data Science and Digital Solutions

A company must retain customer records for seven years and delete them afterward. Which governance artifact should define this rule?

  1. A dashboard style guide
  2. A feature selection log
  3. A model training script
  4. A data retention policy Correct answer.

Explanation

A data retention policy specifies how long each class of data is kept and when it must be disposed of, satisfying legal and regulatory obligations. Style guides, feature logs, and training scripts do not govern record lifecycle.

Question 21Data Science and Digital Solutions

A team wants to reduce 200 correlated features to a smaller set of uncorrelated components before clustering. Which technique is appropriate?

  1. One-hot encoding
  2. Stratified sampling
  3. Principal component analysis Correct answer.
  4. Logistic regression

Explanation

Principal component analysis linearly transforms correlated features into a smaller set of orthogonal components ordered by explained variance. One-hot encoding handles categorical variables, stratified sampling addresses class representation, and logistic regression is a supervised classifier.

Question 22Data Science and Digital Solutions

A model performs well on training data but poorly on held-out data. Which remedy most directly addresses this gap?

  1. Removing the validation set
  2. Increasing model complexity
  3. Training for many more epochs on the same data
  4. Regularization combined with more training data Correct answer.

Explanation

The pattern indicates overfitting, where the model memorizes training noise. Regularization constrains model flexibility and additional data gives it more signal to learn, both improving generalization, whereas more complexity or epochs worsens the gap.

Question 23Data Science and Digital Solutions

An analyst runs 40 independent hypothesis tests at a significance level of 0.05 with no true effects present. Roughly how many false positives should be expected, and what correction applies?

  1. About 40 false positives; reduce the sample size
  2. About 0 false positives; no correction is needed
  3. About 20 false positives; increase the significance level
  4. About 2 false positives; apply a multiple-comparison correction such as Bonferroni Correct answer.

Explanation

With 40 independent tests at 0.05, the expected number of false positives is 40 times 0.05, or about 2. Multiple-comparison corrections such as Bonferroni lower the per-test threshold to control the family-wise error rate.

Question 24Data Science and Digital Solutions

A data engineer needs a processing pattern that handles unbounded event streams with low latency while tolerating late-arriving records. Which design fits best?

  1. A streaming pipeline with event-time windows and watermarks Correct answer.
  2. A static report generated once per quarter
  3. A nightly batch job that reloads the full history
  4. A weekly manual export reviewed by an analyst

Explanation

Streaming systems process events continuously, and event-time windows with watermarks let them wait briefly for out-of-order records while still emitting timely results. Batch, manual, and periodic reporting introduce latency and cannot accommodate late events gracefully.

Question 25Data Science and Digital Solutions

A team trains a model on data from one region and deploys it in another where customer behavior differs sharply. Which issue does this scenario primarily illustrate?

  1. Overfitting to the test set
  2. Data drift between training and deployment distributions Correct answer.
  3. Label leakage within the training set
  4. Multicollinearity among predictors

Explanation

When the deployment population differs from the training population, the input distribution has shifted, a form of data or covariate drift that can degrade performance. Leakage, test-set overfitting, and multicollinearity are different problems arising within the modeling process itself.

Question 26Data Science and Digital Solutions

Which data structure is most appropriate for storing a customer's set of unique product IDs viewed in a session when order does not matter?

  1. A stack
  2. A linked list
  3. A queue
  4. A hash set Correct answer.

Explanation

A hash set stores unique elements and supports fast membership checks without preserving order, which matches the requirement of deduplicated product IDs. Stacks and queues impose order semantics that are irrelevant here, and a linked list allows duplicates and slower lookup.

Question 27Data Science and Digital Solutions

Which SQL clause is used to filter individual rows before any grouping occurs?

  1. HAVING
  2. ORDER BY
  3. GROUP BY
  4. WHERE Correct answer.

Explanation

The WHERE clause evaluates row-level conditions before aggregation, so it removes rows prior to grouping. HAVING filters groups after aggregation, ORDER BY sorts results, and GROUP BY forms the groups themselves.

Question 28Data Science and Digital Solutions

In supervised learning, what distinguishes classification from regression?

  1. Classification uses only text inputs while regression uses only numeric inputs
  2. Classification always uses neural networks while regression always uses linear models
  3. Classification predicts discrete labels while regression predicts continuous values Correct answer.
  4. Classification requires no training data while regression requires labeled data

Explanation

The defining difference is the target variable: classification assigns records to discrete categories, whereas regression estimates a continuous quantity. The input type and model family are independent of this distinction.

Question 29Data Science and Digital Solutions

What does a primary key enforce on a relational table?

  1. That every foreign key in the database references this table
  2. That the table is automatically indexed by every column
  3. Uniqueness and non-null values for the chosen column or columns Correct answer.
  4. That rows are physically stored in sorted order

Explanation

A primary key uniquely identifies each row, so its values must be unique and cannot be null. It does not control foreign keys in other tables, index every column, or dictate physical row order.

Question 30Data Science and Digital Solutions

Which measure of central tendency is most affected by a single extreme outlier?

  1. The mean Correct answer.
  2. The interquartile range
  3. The mode
  4. The median

Explanation

The mean sums all values, so one very large or small observation shifts it substantially. The median depends on rank position and the mode on frequency, both of which resist outliers, while the interquartile range is a spread measure rather than a center.

Question 31Data Science and Digital Solutions

In version control, what is the main purpose of a branch?

  1. To permanently delete earlier commits from the repository history
  2. To automatically merge every commit into production
  3. To compress large binary files before they are committed
  4. To isolate a line of development so changes can be made without affecting the main codebase Correct answer.

Explanation

A branch provides an independent line of work so experimental or feature changes do not disturb the stable main line until merged. It does not erase history, compress files, or trigger automatic deployment.

Question 32Data Science and Digital Solutions

Which practice best protects sensitive personal data displayed in an analytics report?

  1. Aggregating or masking identifiers so individuals cannot be re-identified Correct answer.
  2. Removing only the column headers from the export
  3. Storing the extract on a shared drive with open permissions
  4. Emailing the raw extract to all report readers

Explanation

Aggregation and masking reduce identifiability while preserving analytical value, which is the core of privacy-protective reporting. Broad distribution, open storage, and cosmetic edits leave the sensitive values exposed.

Question 33Data Science and Digital Solutions

What does a data pipeline's extract step accomplish?

  1. It trains the final machine learning model
  2. It retrieves data from source systems for downstream processing Correct answer.
  3. It renders charts for end users
  4. It defines access control policies for analysts

Explanation

Extraction pulls raw data from source systems such as databases or APIs so it can be transformed and loaded. Model training, visualization, and access policy are separate stages or concerns.

Question 34Data Science and Digital Solutions

A retail team wants to group customers by purchasing behavior without any predefined segments. Which technique applies?

  1. A paired t-test on two campaign groups
  2. Linear regression on total spend
  3. Logistic regression on churn labels
  4. K-means clustering on standardized purchase features Correct answer.

Explanation

Clustering is an unsupervised method that discovers natural groupings when no segment labels exist. Regression and hypothesis tests require a defined target or comparison groups, which the scenario lacks.

Question 35Data Science and Digital Solutions

A dataset has 5 percent missing values in a numeric feature that appear to be missing completely at random. Which approach is generally most defensible for a baseline model?

  1. Impute the median and add a missing-indicator flag Correct answer.
  2. Replace missing values with zero and ignore the pattern
  3. Delete every row containing any missing value
  4. Drop the feature entirely from the dataset

Explanation

Median imputation is robust to skew, and an indicator flag preserves the information that a value was absent. Deleting all affected rows discards data, zero-filling distorts the distribution, and dropping the feature loses potentially useful signal.

Question 36Data Science and Digital Solutions

An analyst joins a customers table to an orders table on customer ID. Customers with no orders disappear from the result. Which join preserves all customers?

  1. A left join with customers on the left Correct answer.
  2. A right join with customers on the right
  3. An inner join with customers on the left
  4. A cross join of both tables

Explanation

A left join keeps every row from the left table and fills unmatched right-side columns with nulls, so customers without orders remain. An inner join keeps only matches, a cross join produces a Cartesian product, and a right join would preserve the orders side instead.

Question 37Data Science and Digital Solutions

A team wants to compare conversion rates between two website variants and needs a statistical test for the difference in proportions. Which test is appropriate?

  1. A one-sample t-test on page views
  2. Pearson correlation between variant and revenue
  3. A chi-square goodness-of-fit test on a single variable
  4. A two-proportion z-test Correct answer.

Explanation

Comparing conversion proportions across two independent groups calls for a two-proportion z-test, which models the difference in binomial rates. A one-sample t-test, goodness-of-fit test, and correlation coefficient address different questions.

Question 38Data Science and Digital Solutions

A model's predicted probabilities are all close to 0.5, and the ranking of records by risk is nearly random. Which metric best exposes this weakness?

  1. The mean of the predicted probabilities
  2. The training time per epoch
  3. The area under the ROC curve Correct answer.
  4. The number of features used

Explanation

The area under the ROC curve measures how well scores rank positives above negatives, and a value near 0.5 signals near-random discrimination. The average probability, feature count, and training time say nothing about ranking quality.

Question 39Data Science and Digital Solutions

A weekly report must show how the mix of three product categories changed over twelve months. Which visualization fits best?

  1. A histogram of monthly revenue
  2. A scatter plot of price versus quantity
  3. A pie chart of a single month
  4. A stacked area chart of category share by month Correct answer.

Explanation

A stacked area chart displays part-to-whole composition across a continuous time axis, making shifts in category mix visible. A single-month pie chart, a revenue histogram, and a price-quantity scatter plot cannot show compositional change over time.

Question 40Data Science and Digital Solutions

A team needs a reproducible environment so that a model trained today yields identical results on another machine. Which practice supports this most directly?

  1. Pinning library versions and recording the random seed Correct answer.
  2. Renaming the project folder
  3. Increasing the number of CPU cores
  4. Adding more exploratory notebooks

Explanation

Reproducibility requires controlling both code dependencies and randomness, so pinned versions and a recorded seed let others recreate the same run. Hardware, folder names, and extra notebooks do not fix the sources of variation.

Question 41Data Science and Digital Solutions

A normalized operational database slows down a reporting query that aggregates millions of sales rows. Which design choice most directly improves read performance for this workload?

  1. Build a denormalized analytical table or star schema for reporting Correct answer.
  2. Add more normalization steps to the sales tables
  3. Store the sales data only in flat CSV files
  4. Remove all indexes from the sales tables

Explanation

Analytical workloads benefit from denormalized structures such as star schemas that reduce joins and support fast aggregation. Further normalization adds joins, removing indexes slows reads, and flat files lack query optimization.

Question 42Data Science and Digital Solutions

An analyst wants to estimate the average order value for a population but the sample contains a few extremely large orders. Which interval estimate is most appropriate?

  1. The sample maximum as the upper bound
  2. A confidence interval that assumes a normal sample mean without checking
  3. A bootstrap confidence interval for the mean Correct answer.
  4. The sample median reported without any interval

Explanation

Bootstrapping resamples the observed data to approximate the sampling distribution, which remains reliable when the mean's distribution is distorted by heavy tails. Assuming normality ignores the skew, the maximum is not a bound on the mean, and a bare median gives no uncertainty estimate.

Question 43Data Science and Digital Solutions

A fraud team must explain each declined transaction to regulators. Which model property matters most?

  1. Number of hidden layers
  2. Interpretability of the decision logic Correct answer.
  3. Size of the feature store
  4. Raw training speed

Explanation

Regulatory explanation requires that the reasons for each decision be understandable, so interpretable models or post-hoc explanation methods are essential. Training speed, network depth, and storage size do not provide justifications.

Question 44Data Science and Digital Solutions

A data scientist needs to compare model performance across five algorithms using the same cross-validation folds. Why is using identical folds important?

  1. It guarantees that every algorithm will achieve high accuracy
  2. It removes the need for a test set entirely
  3. It ensures the algorithms use identical hyperparameters
  4. It controls fold-to-fold variation so differences reflect the algorithms rather than the split Correct answer.

Explanation

Sharing folds holds the data partition constant, so performance differences can be attributed to the models instead of sampling noise in the splits. It does not guarantee accuracy, replace final testing, or equalize hyperparameters.

Question 45Data Science and Digital Solutions

A dashboard must refresh within seconds as new sensor readings arrive, and queries must scan only recent data. Which storage pattern best supports this?

  1. Archiving every reading into a PDF report
  2. Recomputing the entire history on each refresh
  3. Partitioning the time-series table by ingestion date Correct answer.
  4. Storing all history in a single unpartitioned table

Explanation

Partitioning by date lets queries prune to recent partitions, cutting scan volume and latency for fresh data. A single unpartitioned table forces full scans, PDF reports are not queryable, and full recomputation is slow and wasteful.

Question 46Data Science and Digital Solutions

A team notices that a deployed model's input feature distributions have drifted far from training data, yet labels are unavailable. Which monitoring approach is most suitable?

  1. Monitor server CPU utilization as a proxy for drift
  2. Track population stability and distribution distances on input features Correct answer.
  3. Retrain weekly regardless of any measured change
  4. Wait for annual accuracy reviews using labeled outcomes

Explanation

When labels are delayed or absent, input-distribution monitoring such as population stability indices or divergence measures detects drift without ground truth. Accuracy reviews need labels, CPU usage is unrelated to data drift, and blind retraining ignores evidence.

Question 47Data Science and Digital Solutions

A pipeline processes events that may arrive out of order and must produce correct aggregates over fixed time windows. Which mechanism is essential?

  1. Processing-time windows that ignore event timestamps
  2. A single global counter with no window boundaries
  3. Buffering all events indefinitely before any output
  4. Event-time windowing with watermarks to bound late arrivals Correct answer.

Explanation

Event-time windows group records by when they occurred, and watermarks define how long the system waits for late data before emitting results. Processing-time windows misassign late events, a global counter loses window semantics, and infinite buffering prevents timely output.

Question 48Data Science and Digital Solutions

A team must prove that a trained model satisfies a fairness constraint across protected groups before release. Which step is most appropriate?

  1. Report only overall accuracy for the full population
  2. Increase the model's parameter count until metrics improve
  3. Remove the protected attribute from the feature set and stop there
  4. Measure group-level error rates and compare them against a predefined threshold Correct answer.

Explanation

Fairness constraints are verified by computing outcome or error metrics per group and checking them against agreed thresholds. Overall accuracy can hide disparities, simply dropping the attribute does not remove proxy effects, and adding parameters is unrelated to the constraint.

Question 49Data Science and Digital Solutions

An organization wants lineage tracking so that any reported figure can be traced back to its source tables and transformations. Which capability is most central?

  1. A metadata catalog recording transformation dependencies Correct answer.
  2. A more permissive access policy
  3. A shorter refresh schedule
  4. A larger data warehouse cluster

Explanation

Lineage depends on metadata that records how each dataset derives from upstream sources and transformations, which a catalog captures. Cluster size, access policy, and refresh cadence affect performance or governance but not traceability.

Question 50Data Science and Digital Solutions

Which statement best describes the difference between supervised and unsupervised learning?

  1. Unsupervised learning can only be applied to text data rather than numeric data.
  2. Supervised learning uses labeled target values, while unsupervised learning looks for structure without labels. Correct answer.
  3. Supervised learning does not require any training data before making predictions.
  4. Supervised learning always requires more computing power than unsupervised learning.

Explanation

The defining feature of supervised learning is that each training example carries a known target label that the model learns to predict. Unsupervised learning has no target label and instead seeks patterns such as clusters or reduced dimensions. Computing cost and data type are not what separates the two paradigms.