Trending Technologies MCQs

Synthetic Data MCQs (Multiple-Choice Questions)

Practice Synthetic Data MCQs to test your knowledge of synthetic data generation, statistical modeling, generative AI, data privacy, and machine learning applications. These questions cover the methods used to create artificial datasets that reproduce useful characteristics of real-world data. They are useful for students, data scientists, AI developers, researchers, and professionals preparing for technical interviews and examinations. The set includes both foundational and practical questions covering modern synthetic data systems.

Synthetic Data MCQs

These Synthetic Data multiple-choice questions cover important concepts such as data fidelity, statistical distributions, Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), diffusion models, differential privacy, data augmentation, and synthetic dataset evaluation. This set combines conceptual, technical, and scenario-based questions to help test your understanding of synthetic data systems.

Synthetic Data MCQs cover the technologies used to generate artificial data, preserve useful patterns, support AI development, and assess privacy risks. Each question includes an answer and explanation.

List of Synthetic Data MCQs

The following Synthetic Data multiple-choice questions cover generation techniques, data quality, privacy protection, machine learning applications, evaluation metrics, and practical implementation scenarios.

1. What is synthetic data?

Data collected exclusively from physical sensors
Data created artificially using algorithms, simulations, or statistical models
Data that contains only handwritten records
Data that cannot be stored electronically

Answer: B) Data created artificially using algorithms, simulations, or statistical models

Explanation:

Synthetic data is artificially generated rather than directly collected from real-world events for each generated record. It can be designed to reproduce selected statistical patterns, relationships, or behaviors of real data for analysis, testing, and machine learning.

2. What is a primary purpose of synthetic data generation?

To eliminate the need for data quality checks
To guarantee perfect machine learning accuracy
To replace all databases with random values
To create useful artificial datasets when real data is limited, sensitive, or difficult to obtain

Answer: D) To create useful artificial datasets when real data is limited, sensitive, or difficult to obtain

Explanation:

Synthetic data can support model development, software testing, research, and simulation when access to real data is restricted or insufficient. Its usefulness depends on whether it preserves the properties required by the intended task.

3. Which of the following is an example of synthetic tabular data?

A generated dataset containing artificial customer ages, purchase amounts, and account categories
A photograph captured directly by a camera
A real-time temperature measurement collected by a physical sensor
An unmodified transcript of an actual customer call

Answer: A) A generated dataset containing artificial customer ages, purchase amounts, and account categories

Explanation:

Synthetic tabular data contains artificially generated rows and columns, often designed to resemble the structure and statistical characteristics of a real dataset. The generated records can be used for analytics, testing, and machine learning when they are suitable for the task.

4. What does data fidelity mean when evaluating synthetic data?

The speed at which a file is downloaded
The number of columns in a spreadsheet
The degree to which synthetic data preserves relevant characteristics of the reference data
The physical size of the storage device

Answer: C) The degree to which synthetic data preserves relevant characteristics of the reference data

Explanation:

Data fidelity measures how closely generated data matches selected properties of the reference data, such as distributions, correlations, or temporal patterns. High fidelity in one aspect does not guarantee good downstream performance or strong privacy protection.

5. Which type of synthetic data is designed to represent sequences of observations recorded over time?

Static metadata only
Time-series synthetic data
Database schema definitions only
Unstructured directory listings

Answer: B) Time-series synthetic data

Explanation:

Time-series synthetic data represents sequential observations such as sensor readings, financial values, or equipment measurements. A suitable generator should preserve relevant temporal behavior, including trends, seasonality, dependencies, and event patterns.

6. Which technique can generate synthetic images for computer vision training?

SQL indexing alone
DNS resolution
HTML minification
Generative image models

Answer: D) Generative image models

Explanation:

Generative image models can create artificial images resembling selected visual categories or conditions. They can support computer vision training, testing, and augmentation, provided the generated images and labels are appropriate for the target task.

7. What is a Generative Adversarial Network (GAN)?

A generative model framework involving a generator and a discriminator trained in competition
A database backup method
A deterministic file compression algorithm
A rule-based system that can only sort existing records

Answer: A) A generative model framework involving a generator and a discriminator trained in competition

Explanation:

A GAN contains a generator that creates synthetic samples and a discriminator that attempts to distinguish generated samples from real ones. Training uses this competition to improve the generator's output, although GANs can experience instability and mode collapse.

8. What is the role of the generator in a GAN?

To delete the original training dataset
To label every real record manually
To produce synthetic samples intended to resemble the training distribution
To evaluate only the storage capacity of a computer

Answer: C) To produce synthetic samples intended to resemble the training distribution

Explanation:

The generator maps input noise or another representation to synthetic samples. During GAN training, it learns to produce samples that are difficult for the discriminator to distinguish from real training examples.

9. What is the primary function of the discriminator in a standard GAN?

To create a database schema
To estimate whether a sample is real or generated
To calculate the physical temperature of a GPU
To convert every synthetic record into a text document

Answer: B) To estimate whether a sample is real or generated

Explanation:

The discriminator learns to distinguish real training samples from synthetic samples produced by the generator. Its feedback contributes to generator training. In conditional GANs, both networks may also use condition information such as class labels.

10. What is mode collapse in GAN training?

When a model produces a wide range of realistic samples
When all training examples are automatically labeled correctly
When the discriminator is removed from every GAN architecture
When the generator produces limited varieties of samples and fails to capture the full data distribution

Answer: D) When the generator produces limited varieties of samples and fails to capture the full data distribution

Explanation:

Mode collapse occurs when a generator produces samples from only a limited subset of possible patterns. The output may look plausible but lack diversity, causing important categories or rare patterns in the reference data to be underrepresented.

11. What is a Variational Autoencoder (VAE)?

A model that learns a probabilistic latent representation and can generate samples by decoding latent variables
A file system that stores duplicated records
A sorting algorithm for categorical columns
A system that only converts images into filenames

Answer: A) A model that learns a probabilistic latent representation and can generate samples by decoding latent variables

Explanation:

A VAE uses an encoder to learn a distribution over latent representations and a decoder to reconstruct or generate samples. After training, latent variables can be sampled and passed through the decoder to produce synthetic data.

12. What is the latent space in a generative model?

A folder containing raw text files only
A network connection used for model deployment
A learned or defined representation space from which meaningful data patterns can be encoded or generated
A list of database passwords

Answer: C) A learned or defined representation space from which meaningful data patterns can be encoded or generated

Explanation:

The latent space represents data in a compact or transformed form. In many generative models, points in this space can be decoded into samples. The structure and usefulness of the latent space depend on the model architecture and training objective.

13. What is a diffusion model commonly trained to do?

Sort database tables by row count
Learn to reverse a gradual noise-corruption process to generate samples
Replace every real value with zero
Generate output without any learned parameters

Answer: B) Learn to reverse a gradual noise-corruption process to generate samples

Explanation:

Diffusion models typically learn to remove or reverse noise introduced into training data through a sequence of steps. Generation starts from noise or another initial distribution and iteratively produces a sample, such as an image or another data representation.

14. Which model family is particularly useful for generating realistic synthetic text?

DNS servers
Spreadsheet formulas alone
Static image filters only
Transformer-based language models

Answer: D) Transformer-based language models

Explanation:

Transformer-based language models can generate text by predicting tokens conditioned on preceding context or other inputs. They can produce synthetic conversations, documents, and structured text, but outputs must be checked for accuracy, representativeness, privacy risks, and unwanted bias.

15. What is conditional synthetic data generation?

Generating data based on specified conditions such as labels, attributes, or scenarios
Generating random data without any control or input
Deleting all features before generation
Generating data only when the computer is disconnected from a network

Answer: A) Generating data based on specified conditions such as labels, attributes, or scenarios

Explanation:

Conditional generation allows users to specify attributes or conditions that influence the generated samples. For example, an image generator may create examples of a particular object category, or a tabular generator may generate records with a chosen target label.

16. What is the main purpose of synthetic data augmentation?

To remove all training examples
To ensure that the test set contains only duplicates
To increase the diversity or quantity of training examples using generated data
To eliminate the need to measure model performance

Answer: C) To increase the diversity or quantity of training examples using generated data

Explanation:

Synthetic data augmentation adds generated examples to a training dataset. It can help expose models to additional variations or underrepresented cases, but the generated samples should be realistic and relevant to the task. Poor-quality augmentation can introduce artifacts or weaken model performance.

17. Why might synthetic data be useful for rare-event modeling?

It guarantees that rare events will never occur in production
It can provide additional examples of uncommon scenarios that are scarce in real datasets
It removes the need to define the target event
It ensures that every rare event is statistically independent

Answer: B) It can provide additional examples of uncommon scenarios that are scarce in real datasets

Explanation:

Rare events such as equipment failures or fraudulent transactions may be poorly represented in collected data. Synthetic examples can help train or test systems for these cases, provided the generated patterns reflect plausible scenarios rather than introducing misleading correlations.

18. What does class imbalance mean in a machine learning dataset?

All classes have exactly the same number of samples
The dataset contains no target labels
Every record has an identical feature vector
Some classes have substantially more examples than others

Answer: D) Some classes have substantially more examples than others

Explanation:

Class imbalance occurs when target categories have unequal sample counts. A model trained on an imbalanced dataset may perform poorly on minority classes. Carefully generated synthetic minority examples can help, but evaluation should use an appropriate held-out dataset that reflects the intended deployment setting.

19. How can synthetic data help software testing?

By supplying controlled test records without requiring actual customer records for every scenario
By guaranteeing that software contains no defects
By automatically replacing all source code
By preventing developers from testing edge cases

Answer: A) By supplying controlled test records without requiring actual customer records for every scenario

Explanation:

Synthetic records can represent normal inputs, edge cases, boundary values, and unusual scenarios in development and testing environments. They help teams test behavior without depending entirely on production data, although the generated cases must still reflect relevant requirements and failure conditions.

20. What is a key benefit of synthetic data in autonomous vehicle development?

It eliminates the need to test vehicles in real conditions
It guarantees safe driving in every possible environment
It enables simulation of varied driving scenarios, including dangerous or rare situations
It makes sensor calibration unnecessary

Answer: C) It enables simulation of varied driving scenarios, including dangerous or rare situations

Explanation:

Simulated environments can generate diverse road layouts, weather conditions, traffic situations, and rare hazards. This supports repeatable testing and data generation. However, simulated data may not capture every real-world detail, so validation with realistic scenarios and real-world evidence remains necessary.

21. What is differential privacy?

A method for increasing the resolution of synthetic images
A mathematical framework that bounds how much a computation's output can reveal about an individual's data
A database format used only for numerical data
A technique that guarantees every generated record is identical to a real record

Answer: B) A mathematical framework that bounds how much a computation's output can reveal about an individual's data

Explanation:

Differential privacy provides a formal framework for limiting the influence of an individual record on a computation's output. A synthetic dataset generated with a properly implemented differentially private mechanism can offer a quantified privacy guarantee, depending on the mechanism and its privacy parameters.

22. Does generating synthetic data automatically guarantee privacy?

Yes, because generated data can never resemble real records
Yes, provided the dataset contains more than 1,000 rows
Yes, if the original data is stored in a spreadsheet
No, synthetic data may memorize or reveal information from the training data unless privacy risks are assessed and controlled

Answer: D) No, synthetic data may memorize or reveal information from the training data unless privacy risks are assessed and controlled

Explanation:

Some generators can reproduce rare or near-identical training records, creating disclosure risks. Synthetic data is not automatically anonymous. Privacy assessment, access controls, appropriate generation methods, and differential privacy where suitable can help reduce risks.

23. What is a membership inference attack against a generative model?

An attempt to determine whether a particular record was included in the model's training data
An attempt to increase the size of a generated dataset
A process for calculating the mean of a column
A technique for creating a new database table

Answer: A) An attempt to determine whether a particular record was included in the model's training data

Explanation:

A membership inference attack attempts to infer whether a specific example was used to train a model. Generative models may leak information about training records under some conditions, so membership inference testing can form part of a broader privacy risk assessment.

24. What is the purpose of a privacy budget in differential privacy?

To specify the maximum size of a dataset in megabytes
To determine how many columns a table can contain
To quantify or track privacy loss associated with differentially private operations
To set the price of a data storage subscription

Answer: C) To quantify or track privacy loss associated with differentially private operations

Explanation:

Differential privacy uses parameters such as epsilon and, in approximate variants, delta to express privacy guarantees. Repeated operations can consume privacy budget through composition. Smaller epsilon generally indicates stronger privacy protection for comparable settings, but utility and the complete mechanism must also be considered.

25. What does the parameter epsilon (ε) commonly represent in differential privacy?

The number of records in a generated dataset
A privacy-loss parameter that controls the strength of a differential privacy guarantee
The sampling rate of an audio recording
The number of model layers in a neural network

Answer: B) A privacy-loss parameter that controls the strength of a differential privacy guarantee

Explanation:

Epsilon bounds how much the output distributions of a differentially private mechanism can differ when one individual's data is changed or removed, according to the chosen definition. Lower epsilon usually corresponds to stronger privacy, often at the cost of less accurate results.

26. What is the purpose of statistical distribution comparison when evaluating synthetic data?

To determine the physical location of the generator
To guarantee that no generated value is repeated
To count only the number of file formats used
To assess whether important variables have similar distributions in the synthetic and reference datasets

Answer: D) To assess whether important variables have similar distributions in the synthetic and reference datasets

Explanation:

Comparing distributions helps identify differences in central tendency, spread, category frequencies, and other statistical properties. Similar marginal distributions alone do not prove that relationships between variables are preserved, so additional evaluation is often needed.

27. Why is preserving correlations between variables important in synthetic tabular data?

It helps maintain meaningful relationships among features that downstream analyses may depend on
It guarantees that all records are unique
It eliminates the need to generate categorical values
It ensures that the dataset contains no missing values

Answer: A) It helps maintain meaningful relationships among features that downstream analyses may depend on

Explanation:

Real datasets often contain relationships between variables, such as income and spending or temperature and energy demand. If a generator preserves individual feature distributions but destroys important dependencies, analyses and predictive models using the synthetic data may produce misleading results.

28. What is the train-on-synthetic, test-on-real (TSTR) evaluation approach?

Training and testing only on the same synthetic records
Testing a real-world system without any training
Training a predictive model on synthetic data and evaluating it on held-out real data
Training on real data and testing exclusively on synthetic data

Answer: C) Training a predictive model on synthetic data and evaluating it on held-out real data

Explanation:

TSTR measures whether synthetic data supports a downstream task by training a model on synthetic examples and testing it against real held-out examples. Results can indicate practical utility, but they depend on the task, test set, evaluation metric, and model choice.

29. What is the purpose of a train-on-real, test-on-synthetic (TRTS) evaluation?

To measure network bandwidth
To assess model performance on synthetic test data after training on real data
To guarantee differential privacy
To verify that every generated record is unique

Answer: B) To assess model performance on synthetic test data after training on real data

Explanation:

TRTS trains a model on real data and evaluates it on synthetic data. It can help assess whether the synthetic dataset behaves similarly for a particular predictive task. It should be interpreted alongside other evaluations because good TRTS performance does not establish privacy or full statistical fidelity.

30. What is a potential weakness of evaluating synthetic data using only one statistical metric?

One metric always measures every possible property
A single metric automatically measures privacy and fairness
Statistical metrics cannot be calculated for synthetic data
It may miss important differences in utility, diversity, privacy, or subgroup behavior

Answer: D) It may miss important differences in utility, diversity, privacy, or subgroup behavior

Explanation:

Synthetic data quality is multidimensional. A dataset may closely match average values but fail to preserve rare events, correlations, or minority groups. A robust evaluation should use multiple measures chosen according to the intended use and relevant privacy requirements.

31. What is mode coverage in synthetic data generation?

The extent to which a generator represents the different patterns or subgroups in the target data distribution
The number of times a file is opened
The maximum number of users allowed in a database
The time needed to install a generative model

Answer: A) The extent to which a generator represents the different patterns or subgroups in the target data distribution

Explanation:

Mode coverage concerns whether a generator represents the range of patterns found in the reference distribution. Poor coverage can cause the model to omit important categories, subgroups, or unusual but meaningful cases even if some generated samples appear realistic.

32. What is overfitting in synthetic data generation?

Generating samples that cover all possible real-world situations
Training a model with unlimited computing resources
Learning training-specific details too closely, which can reduce generalization and increase memorization risk
Creating a dataset with a balanced class distribution

Answer: C) Learning training-specific details too closely, which can reduce generalization and increase memorization risk

Explanation:

An overfitted generator may reproduce particular training examples or fail to capture the broader data distribution. This can reduce sample diversity and increase privacy concerns. Suitable validation, regularization, and privacy analysis can help identify such problems.

33. Why is a held-out validation dataset useful when training a synthetic data generator?

It increases the size of every generated sample
It helps assess generalization and select model settings without relying exclusively on training performance
It removes the need for any privacy assessment
It guarantees that the model will work on every domain

Answer: B) It helps assess generalization and select model settings without relying exclusively on training performance

Explanation:

A validation dataset provides information about performance on examples not used directly to fit model parameters. It can help guide model selection and detect overfitting. Data splitting must be designed carefully to prevent leakage, particularly when records are related or time-dependent.

34. What is data leakage in machine learning?

When a synthetic dataset contains numerical columns
When a generator produces multiple records
When data is stored in more than one file
When information unavailable at prediction time improperly influences training or evaluation

Answer: D) When information unavailable at prediction time improperly influences training or evaluation

Explanation:

Data leakage occurs when information from a validation or test set, or from future events, improperly influences training or model selection. Synthetic data pipelines can also leak information if they use test records to fit a generator and then use generated samples in model training.

35. How can synthetic data support fairness evaluation?

By enabling controlled experiments with selected groups, scenarios, or feature combinations
By automatically removing every form of bias
By guaranteeing identical model outcomes for all individuals
By making subgroup evaluation unnecessary

Answer: A) By enabling controlled experiments with selected groups, scenarios, or feature combinations

Explanation:

Synthetic data can be used to create controlled scenarios and investigate model behavior across groups or conditions. However, the generator may reproduce or introduce biases, so results should be compared with reliable reference data and appropriate fairness measures wherever possible.

36. What is a major challenge when generating synthetic healthcare data?

Medical records contain no relationships between variables
Healthcare data cannot be represented in tables
Preserving clinically meaningful relationships while managing privacy and subgroup representation
All patient records have identical distributions

Answer: C) Preserving clinically meaningful relationships while managing privacy and subgroup representation

Explanation:

Healthcare datasets may contain complex relationships among diagnoses, treatments, demographics, and outcomes. Synthetic records must preserve relevant patterns without disclosing sensitive information or misrepresenting patient subgroups. They require careful validation before being used in research or model development.

37. Which technique is commonly used to generate synthetic tabular data?

DNS caching
Copula-based statistical modeling
HTML rendering
Video transcoding alone

Answer: B) Copula-based statistical modeling

Explanation:

Copula-based methods model relationships among variables by connecting marginal distributions through a dependence structure. They can be useful for tabular data generation, although their effectiveness depends on the data's characteristics and the suitability of the chosen model.

38. What is the Synthetic Data Vault (SDV) used for?

Managing physical server cooling systems
Compressing audio recordings
Designing network routing protocols
Providing tools and models for generating and evaluating synthetic data

Answer: D) Providing tools and models for generating and evaluating synthetic data

Explanation:

SDV is a software ecosystem for synthetic data generation, particularly for structured data such as tables and relational datasets. Its tools support different generation approaches and workflows. Users still need to select appropriate models and assess utility, fidelity, and privacy for their use cases.

39. Why is schema preservation important when generating synthetic database records?

It ensures that generated records conform to required fields, data types, and structural constraints
It guarantees that every value is statistically identical to the original
It removes the need for validation rules
It ensures that a database contains no duplicate values

Answer: A) It ensures that generated records conform to required fields, data types, and structural constraints

Explanation:

Schema preservation ensures that synthetic records follow the expected data structure, including columns, types, relationships, and constraints. A dataset can match the schema but still have unrealistic values or incorrect statistical relationships, so structural validation is only one part of quality assessment.

40. What is referential integrity in a relational synthetic dataset?

Ensuring that every value in every column is unique
Ensuring that all tables contain the same number of rows
Ensuring that relationships such as foreign keys point to valid corresponding records
Ensuring that all numeric values are positive

Answer: C) Ensuring that relationships such as foreign keys point to valid corresponding records

Explanation:

Referential integrity maintains valid relationships between related tables. For example, an order record should reference an existing customer record when the schema requires it. Synthetic relational datasets need to preserve these relationships to support realistic application testing and analysis.

41. What is a potential limitation of synthetic data for training a machine learning model?

It cannot be stored in digital form
It may fail to represent important real-world patterns or contain artifacts introduced by the generator
It always requires more storage than real data
It cannot contain labels or target variables

Answer: B) It may fail to represent important real-world patterns or contain artifacts introduced by the generator

Explanation:

A synthetic dataset can be large but still unrepresentative. If its distributions, labels, correlations, or edge cases differ from real deployment data, a model trained on it may generalize poorly. Evaluating downstream performance on appropriate real-world data is important.

42. What is a synthetic-to-real domain gap?

The difference between two file names
The difference between the number of columns in two spreadsheets
The number of synthetic records generated per minute
The mismatch between the distributions or characteristics of synthetic data and real-world data

Answer: D) The mismatch between the distributions or characteristics of synthetic data and real-world data

Explanation:

A synthetic-to-real domain gap occurs when generated samples differ from the conditions encountered in actual use. In computer vision, for example, simulated lighting or textures may not fully match real camera images. Domain adaptation and evaluation on representative real data can help identify and address this gap.

43. Why should generated synthetic data be checked for duplicate or near-duplicate records?

To identify possible memorization, low diversity, or redundant samples
To guarantee that all statistical distributions are correct
To eliminate the need for privacy testing
To ensure that every record contains the same information

Answer: A) To identify possible memorization, low diversity, or redundant samples

Explanation:

Duplicate and near-duplicate checks can reveal low sample diversity or possible reproduction of training records. Such findings do not prove that a model has violated privacy, but they can indicate a need for further investigation into memorization, data utility, and disclosure risk.

44. What is a useful practice when preparing synthetic data for public release?

Assume that artificial generation removes every legal obligation
Publish all training records alongside the generated data
Assess disclosure risk, validate intended utility, and document generation methods and limitations
Skip all testing because the records are not directly collected

Answer: C) Assess disclosure risk, validate intended utility, and document generation methods and limitations

Explanation:

Public release should be preceded by appropriate privacy risk assessment and quality evaluation. Documentation should describe how the data was generated, its intended uses, known limitations, and relevant privacy measures. Synthetic data may still reproduce sensitive information or be unsuitable for particular uses.

45. What is the purpose of scenario-based synthetic data generation?

To generate only average cases and ignore unusual situations
To create records or simulations representing specified conditions and events
To ensure that every generated sample is identical
To remove all control over the generation process

Answer: B) To create records or simulations representing specified conditions and events

Explanation:

Scenario-based generation creates data for selected situations, such as equipment failure, heavy traffic, fraud attempts, or unusual weather. It is valuable for testing systems under controlled conditions, but the scenarios and their probabilities should be realistic for the intended application.

46. How can synthetic data help develop models for autonomous systems?

By replacing every physical sensor with a spreadsheet
By guaranteeing that simulated systems never fail
By removing the need for system-level testing
By generating diverse simulated sensor inputs and edge cases for training and evaluation

Answer: D) By generating diverse simulated sensor inputs and edge cases for training and evaluation

Explanation:

Synthetic sensor data can represent different operating conditions and rare scenarios that are difficult to capture in real-world collection. Its value depends on how faithfully the simulation represents relevant sensor behavior and environmental conditions. Real-world validation remains necessary.

47. Which approach is most appropriate when synthetic data will be used to evaluate a predictive model intended for real-world deployment?

Evaluate only on the same synthetic records used to train the model
Assume synthetic data is always representative of production data
Use an independent, representative real-world test set where available and assess relevant performance metrics
Choose the model with the largest number of generated records without testing accuracy

Answer: C) Use an independent, representative real-world test set where available and assess relevant performance metrics

Explanation:

Real-world evaluation helps determine whether a model trained or developed with synthetic data generalizes to the intended environment. The test set should be independent and representative, with metrics selected for the task. Privacy, fairness, and subgroup performance may require additional assessments.

48. What is the main trade-off when generating differentially private synthetic data?

Balancing the strength of privacy protection against the accuracy and usefulness of generated data
Choosing between an HTML file and a PDF file
Deciding whether a database should have a name
Balancing image resolution against monitor brightness only

Answer: A) Balancing the strength of privacy protection against the accuracy and usefulness of generated data

Explanation:

Differential privacy introduces a formal privacy guarantee, but the mechanisms used can reduce statistical accuracy or downstream utility. The trade-off depends on the privacy parameters, data distribution, generation method, and intended analysis. The appropriate balance should be selected according to the risks and requirements of the application.

49. A bank wants to generate synthetic transaction records to test fraud detection software without exposing actual customer transactions. What should the team do first?

Generate random transaction amounts without defining any testing requirements
Define the required transaction patterns, fraud scenarios, data constraints, and privacy requirements before selecting a generator
Copy real transactions and change only the customer names
Use only normal transactions and exclude all fraud cases

Answer: B) Define the required transaction patterns, fraud scenarios, data constraints, and privacy requirements before selecting a generator

Explanation:

The team should identify which patterns matter for testing, including legitimate transactions, fraud scenarios, class imbalance, temporal behavior, and relationships among fields. It should then select a suitable generation method, validate the generated data, and assess privacy risks. Replacing names alone does not guarantee that real transactions are adequately protected.

50. A healthcare research team generates synthetic patient records for training a disease prediction model. The synthetic dataset closely matches average patient ages and diagnosis frequencies, but the model performs poorly on a held-out real patient dataset. What is the best next step?

Assume the model is ready because the synthetic dataset matches two summary statistics
Increase the number of synthetic records without investigating their quality
Remove all minority patient groups to simplify the dataset
Investigate feature relationships, subgroup coverage, label quality, and synthetic-to-real differences, then reevaluate the generator and downstream model

Answer: D) Investigate feature relationships, subgroup coverage, label quality, and synthetic-to-real differences, then reevaluate the generator and downstream model

Explanation:

Matching average ages and diagnosis frequencies does not guarantee that the synthetic data preserves clinically meaningful relationships or supports accurate predictions. The team should examine feature dependencies, subgroup representation, label quality, and the synthetic-to-real domain gap. It should improve the generation process where needed and evaluate the final model on independent, representative real-world data. Privacy and fairness should also be assessed before clinical use.

Comments and Discussions!

Load comments ↻

Synthetic Data MCQs (Multiple-Choice Questions)