Master of Data Science Exam — Questions and Answers

Question 1: What is the interpretation of a 95% prediction interval for a new observation in linear regression?

95% of the sample data falls within this interval
The mean response at that predictor value is within this interval 95% of the time
The true regression line falls within this interval 95% of the time
A new individual observation will fall within this interval with approximately 95% probability (Correct answer)

Correct answer: A new individual observation will fall within this interval with approximately 95% probability

A 95% prediction interval accounts for both estimation uncertainty and individual variation, capturing where a single new observation is expected to fall.

Question 2: Which cross-validation strategy is most appropriate when working with time-series data in a predictive modeling research project?

Time-series split (expanding window) cross-validation (Correct answer)
Leave-one-out cross-validation
Stratified k-fold cross-validation
Standard k-fold cross-validation

Correct answer: Time-series split (expanding window) cross-validation

Time-series split preserves the temporal ordering of observations, ensuring the model is always trained on past data and tested on future data.

Question 3: The evimp function in the ______ package is wrapped in varImp.

plot
earth (Correct answer)
None of the above
numpy

Correct answer: earth

The `evimp` function is specifically used to calculate variable importance for models generated by the `earth` package, which implements Multivariate Adaptive Regression Splines (MARS). The `varImp` function in the `caret` package provides a unified interface for variable importance across various models, and it wraps `evimp` when dealing with `earth` models.

Question 4: Which of the following focuses on the data's (previously unidentified) qualities being discovered?

Data wrangling
Data mining (Correct answer)
Machine Learning
Big Data

Correct answer: Data mining

Data mining is the process of discovering patterns, insights, and knowledge from large datasets, often previously unidentified. It involves using various techniques from statistics, machine learning, and database systems to extract valuable information. The core objective is to uncover hidden relationships and predictive models within the data.

Question 5: Which of the following supervised learning algorithms is inherently non-parametric and makes predictions for a new data point based on the majority class or average value of its 'k' closest neighbors in the feature space?

K-Nearest Neighbors (KNN) (Correct answer)
Linear Discriminant Analysis (LDA)
Support Vector Machine (SVM)
Logistic Regression

Correct answer: K-Nearest Neighbors (KNN)

K-Nearest Neighbors (KNN) is a non-parametric, instance-based learning algorithm. It does not learn a discriminative function from the training data but instead stores the entire dataset. For a new, unseen data point, it finds the 'k' most similar instances (neighbors) in the training data and predicts the output based on their properties. For classification, it uses a majority vote; for regression, it uses the average of the values. [11]

Question 6: Which Python library provides the DataFrame abstraction and is the standard tool for tabular data manipulation in data science?

scikit-learn
NumPy
pandas (Correct answer)
SciPy

Correct answer: pandas

pandas provides the DataFrame and Series objects with powerful data manipulation, aggregation, and I/O capabilities for tabular data.

Question 7: Which regularization technique randomly deactivates a fraction of neurons during each training step to reduce overfitting?

L2 regularization
Early stopping
Dropout (Correct answer)
Data augmentation

Correct answer: Dropout

Dropout randomly zeros neuron activations with probability p during training, forcing the network to learn redundant representations and reducing co-adaptation.

Question 8: Which assumption is NOT required for the classical ordinary least squares (OLS) estimator to be BLUE (Best Linear Unbiased Estimator)?

Homoscedasticity (constant error variance)
No perfect multicollinearity
Normality of errors (Correct answer)
Linearity in parameters

Correct answer: Normality of errors

The Gauss-Markov theorem guarantees OLS is BLUE under linearity, exogeneity, homoscedasticity, and no perfect multicollinearity — normality is not required.

Question 9: Which statistical measure is MOST resistant to the influence of outliers when describing the center of a distribution?

Median (Correct answer)
Trimmed mean
Mean
Weighted mean

Correct answer: Median

The median is not affected by extreme values because it depends only on the middle rank, not the magnitude of any values.

Question 10: What is the primary purpose of the kernel trick in Support Vector Machines?

To reduce the number of support vectors
To normalize feature scales before classification
To select the optimal regularization parameter C
To map data into a higher-dimensional space without computing the transformation explicitly (Correct answer)

Correct answer: To map data into a higher-dimensional space without computing the transformation explicitly

The kernel trick allows SVMs to operate in a high-dimensional feature space by computing dot products in that space implicitly, using a kernel function. This makes it computationally feasible to find non-linear decision boundaries without explicitly computing the high-dimensional mapping.

Question 11: Which algorithm is commonly used to segment text into subword units, helping models handle out-of-vocabulary words?

Byte Pair Encoding (BPE) (Correct answer)
Viterbi decoding
Latent Dirichlet Allocation
TF-IDF weighting

Correct answer: Byte Pair Encoding (BPE)

Byte Pair Encoding iteratively merges the most frequent character pairs to build a subword vocabulary, allowing models to represent rare or unseen words as combinations of known subword tokens.

Question 12: When applying label encoding to an ordinal categorical feature with categories ['Low', 'Medium', 'High'], which concern is most important to address?

Label encoding cannot be applied to features with more than two categories
Label encoding requires the feature to be normally distributed first
The assigned integer values must preserve the natural order of the categories (Correct answer)
Label encoding always introduces multicollinearity

Correct answer: The assigned integer values must preserve the natural order of the categories

For ordinal features, the integer labels assigned by encoding must reflect the correct ordering (e.g., Low=0, Medium=1, High=2). Using arbitrary or reversed numeric assignments would mislead models that interpret numeric magnitude as meaningful.

Question 13: Which transformation is most appropriate to stabilize variance (correct heteroscedasticity) when the residual spread increases proportionally with the fitted values?

Square root transformation of the predictor
Adding a quadratic term for the predictor
Standardizing all predictors to z-scores
Log transformation of the response variable (Correct answer)

Correct answer: Log transformation of the response variable

When variance grows proportionally with the mean (a common pattern in count or ratio data), a log transformation of the response variable compresses large values more than small ones, stabilizing variance and often linearizing a multiplicative relationship.

Question 14: A data science researcher is choosing between Ridge and Lasso regression for a high-dimensional dataset. What is the key distinguishing behavior of Lasso?

It requires normally distributed predictors
It penalizes the sum of squared coefficients
It can shrink coefficients exactly to zero, performing feature selection (Correct answer)
It always produces lower bias than Ridge

Correct answer: It can shrink coefficients exactly to zero, performing feature selection

Lasso uses L1 regularization which can drive coefficients to exactly zero, effectively selecting a subset of features.

Question 15: A researcher fits a multiple regression model and finds the F-statistic is highly significant (p < 0.001), yet none of the individual t-tests for the coefficients are significant. What is the most likely explanation?

The researcher should use logistic regression instead
Severe multicollinearity is inflating the standard errors of individual coefficients (Correct answer)
The response variable violates the normality assumption
The model has too few observations to detect any effect

Correct answer: Severe multicollinearity is inflating the standard errors of individual coefficients

When the overall F-test is significant but individual t-tests are not, severe multicollinearity is the classic culprit. Highly correlated predictors inflate standard errors, making it difficult to isolate each variable's unique contribution, even though collectively they explain significant variance.

Question 16: What is the primary objective of the Self-Organizing Map (SOM) algorithm?

To partition data into exactly K non-overlapping clusters
To project high-dimensional data onto a low-dimensional grid while preserving topological relationships (Correct answer)
To maximize the variance of projected data along orthogonal axes
To build a probabilistic generative model of the data

Correct answer: To project high-dimensional data onto a low-dimensional grid while preserving topological relationships

SOMs use competitive learning to map high-dimensional input data onto a typically two-dimensional grid, preserving the topological structure so that similar inputs map to nearby neurons.

Question 17: What transformation is typically applied to convert a cyclical feature such as 'hour of day' into usable numerical features?

Min-max scaling the hour to [0, 1]
Label encoding the hour values 0–23
Applying sine and cosine transformations based on the cycle length (Correct answer)
Converting hour to a categorical ordinal variable

Correct answer: Applying sine and cosine transformations based on the cycle length

Sine and cosine transformations map cyclical features onto a circle, ensuring that the model perceives hour 23 and hour 0 as close to each other rather than far apart.

Question 18: The problems with big data veracity go beyond volume, diversity, and velocity.

FALSE
True (Correct answer)

Correct answer: True

The '3 V's' (Volume, Velocity, Variety) are commonly associated with Big Data, but 'Veracity' is a fourth crucial dimension. Veracity refers to the trustworthiness, accuracy, and quality of the data, addressing issues like bias, noise, and abnormalities that can significantly impact analysis and decision-making.

Question 19: Which D3.js concept binds data values to DOM elements and creates, updates, or removes elements as data changes?

Voronoi tessellation
Force simulation
Projection
Data join (enter-update-exit pattern) (Correct answer)

Correct answer: Data join (enter-update-exit pattern)

D3's data join pattern uses .data() to bind arrays, .enter() for new elements, .exit() for removed elements, and the update selection for existing ones.

Question 20: In a multi-class classification problem with severe class imbalance, which averaging strategy for F1-score best accounts for class frequency when reporting overall model performance?

Binary-average
Micro-average
Weighted-average (Correct answer)
Macro-average

Correct answer: Weighted-average

Weighted-average F1-score calculates F1 for each class independently and then takes the average weighted by the number of true instances per class, making it sensitive to class imbalance and giving more influence to larger classes.

Question 21: A p-value of 0.001 in a study with n = 10,000 primarily reflects:

A very small effect size that is detectable only due to the huge sample (Correct answer)
A large effect size
High posterior probability that H₀ is false
An error in the statistical analysis

Correct answer: A very small effect size that is detectable only due to the huge sample

With n = 10,000, even tiny effects produce very small p-values; the p-value conflates effect size with sample size.

Question 22: In the context of Apache Kafka, what is the term for an ordered, immutable sequence of records that is continually appended to?

Log (Correct answer)
Partition
Queue
Topic

Correct answer: Log

Kafka's core abstraction is the log — an ordered, immutable sequence of records. Topics are categories, and partitions are ordered subdivisions of a topic, but the fundamental data structure is the log.

Question 23: A Kafka consumer group with 4 consumers subscribing to a topic with 3 partitions will result in:

An error due to mismatch
All 4 consumers reading all partitions
2 consumers sharing each partition
3 consumers active and 1 idle (Correct answer)

Correct answer: 3 consumers active and 1 idle

Kafka assigns at most one consumer per partition within a group, so one consumer will be idle when consumers outnumber partitions.

Question 24: What does the McNemar test evaluate in model comparison?

Whether the training and test distributions are the same
Whether two models have equal mean squared error
Whether a model's precision and recall are significantly different
Whether the disagreements between two classifiers are statistically symmetric (Correct answer)

Correct answer: Whether the disagreements between two classifiers are statistically symmetric

McNemar's test uses a 2x2 contingency table of cases where classifiers disagree to test if one model significantly outperforms the other.

Question 25: One of the fundamental abilities in data science is which of the following?

Data Visualization (Correct answer)
All of the above
Statistics
Machine Learning

Correct answer: Data Visualization

Data visualization is a fundamental skill in data science because it enables practitioners to explore datasets, identify patterns, communicate findings effectively, and present complex information in an understandable format. It's often the first step in understanding new data and a critical component of explaining results to stakeholders. While machine learning and statistics are also crucial, visualization is key for both exploration and communication.

Question 26: In named entity recognition (NER), which label would typically be assigned to the word 'Stanford' in the sentence 'She studied at Stanford'?

Organization (Correct answer)
Person
Location
Miscellaneous

Correct answer: Organization

Stanford refers to Stanford University, which is classified as an organization entity in standard NER tagging schemes.

Question 27: Which technique helps identify multicollinearity among features during exploratory analysis?

QQ plot
Correlation heatmap (Correct answer)
Histogram
Run chart

Correct answer: Correlation heatmap

A correlation heatmap visualizes pairwise correlations between all features, making it easy to spot highly correlated variable pairs.

Question 28: Which metric is most informative when evaluating a classifier on a highly imbalanced dataset where false negatives are costly?

Accuracy
Specificity
Cohen's Kappa
Recall (Sensitivity) (Correct answer)

Correct answer: Recall (Sensitivity)

Recall measures the proportion of actual positives correctly identified; when false negatives are costly and classes are imbalanced, accuracy is misleading and recall directly captures missed positives.

Question 29: In the Lambda architecture for big data, what is the purpose of the 'speed layer'?

To provide low-latency views of recent data that the batch layer has not yet processed (Correct answer)
To store immutable raw data indefinitely
To manage schema evolution across data stores
To reprocess all historical data in batch

Correct answer: To provide low-latency views of recent data that the batch layer has not yet processed

The speed layer handles real-time data streams to fill the latency gap left by the batch layer's processing delay.

Question 30: Expected Calibration Error (ECE) measures:

The variance of predictions across bootstrap samples
The gap between a model's confidence and its actual accuracy (Correct answer)
The mean absolute error of probability predictions
The difference between training loss and validation loss

Correct answer: The gap between a model's confidence and its actual accuracy

ECE quantifies how well a model's predicted probabilities align with empirical frequencies, with lower ECE indicating better-calibrated predictions.

Question 31: Cook's distance for observation i is used to measure:

The overall influence of observation i on all fitted values (Correct answer)
The correlation between observation i and the predicted response
The leverage of observation i in predictor space
The standardized residual of observation i

Correct answer: The overall influence of observation i on all fitted values

Cook's distance combines leverage and residual size to quantify how much the regression estimates change if observation i is removed.

Question 32: Which technique is used to assess the stability of a regression model by partitioning data into training and validation subsets multiple times?

K-fold cross-validation (Correct answer)
Bootstrap sampling
Jackknife resampling
Stratified random sampling

Correct answer: K-fold cross-validation

K-fold cross-validation splits data into k subsets and iteratively trains/validates across all folds to estimate model performance.

Question 33: A significant drawback of the K-Means clustering algorithm is its sensitivity to the initial placement of cluster centroids. A poor initialization can lead to a suboptimal clustering solution. Which of the following is a common method used to mitigate this issue?

Running the algorithm multiple times with different random initializations and choosing the best result. (Correct answer)
Scaling the data using Min-Max normalization before clustering.
Applying the elbow method to determine the optimal K.
Using a dendrogram to visualize the clusters.

Correct answer: Running the algorithm multiple times with different random initializations and choosing the best result.

Because K-Means can converge to a local optimum depending on the random starting positions of its centroids, a common and effective strategy is to run the algorithm multiple times (n_init) with different random initializations. The final result is then chosen from the run that yields the best performance, typically measured by the lowest within-cluster sum of squares (inertia). The elbow method helps find K but doesn't solve the initialization problem. Scaling is good practice but doesn't directly address initialization sensitivity. Dendrograms are associated with hierarchical clustering.

Question 34: Which metric is most appropriate for evaluating a model when the cost of false negatives is significantly higher than false positives?

Accuracy
Recall (Correct answer)
Specificity
Precision

Correct answer: Recall

Recall measures the proportion of actual positives correctly identified, making it the priority when missing positive cases is costly.

Question 35: Which activation function is most commonly used in hidden layers of deep neural networks to mitigate the vanishing gradient problem?

Sigmoid
ReLU (Correct answer)
Tanh
Softmax

Correct answer: ReLU

ReLU (Rectified Linear Unit) avoids vanishing gradients because its gradient is 1 for positive inputs, enabling deeper networks to train effectively.

Question 36: Which assumption is tested by the Breusch-Pagan test in a regression model?

Independence of errors
Homoscedasticity of residuals (Correct answer)
Normality of residuals
Linearity of the relationship

Correct answer: Homoscedasticity of residuals

The Breusch-Pagan test checks whether the variance of residuals is constant across levels of the fitted values (homoscedasticity).

Question 37: You are working with a customer dataset that includes a 'Subscription Level' feature with the categories: 'Bronze', 'Silver', 'Gold', and 'Platinum'. There is a clear hierarchical order to these categories. Which encoding method would be most suitable and efficient for this feature before using it in a machine learning model?

Binary Encoding
Label Encoding (Correct answer)
Hashing Encoder
One-Hot Encoding

Correct answer: Label Encoding

Label Encoding is ideal for ordinal categorical data where there is an intrinsic, meaningful order among the categories ('Bronze' < 'Silver' < 'Gold' < 'Platinum'). [2] It assigns a unique integer to each category (e.g., 0, 1, 2, 3), preserving the ordinal relationship while being memory-efficient. [3] One-Hot Encoding would be more appropriate for nominal data (no inherent order) but would create unnecessary extra columns and lose the ordinal information. [20]

Question 38: What is the 'vanishing gradient' problem in deep neural networks?

Gradients shrinking exponentially as they propagate back through many layers (Correct answer)
Loss oscillating around a local minimum
Weights not updating due to zero learning rate
Gradients becoming too large during backpropagation

Correct answer: Gradients shrinking exponentially as they propagate back through many layers

In deep networks, repeated multiplication of small gradient values during backpropagation causes them to shrink toward zero, preventing early layers from learning.

Question 39: Which of the following features can be utilized to enhance a classification model's accuracy?

Dependency Grammar
Frequency count of terms
Vector Notation of sentence
All of these (Correct answer)

Correct answer: All of these

All listed features contribute significantly to enhancing a classification model's accuracy in text analysis. Dependency grammar provides structural insights into sentence meaning, while frequency counts of terms (like TF-IDF) capture the importance of words. Vector notation of sentences (e.g., word embeddings) represents semantic relationships, allowing models to understand context and similarity more effectively.

Question 40: A data journalist uses a dual-axis line chart to show revenue (left axis) and customer satisfaction score (right axis) on the same plot. What is the primary risk?

The chart will not render correctly in web browsers
Dual axes require a logarithmic scale
Line charts cannot share an x-axis
Manually chosen axis scales can create misleading apparent correlations or divergences (Correct answer)

Correct answer: Manually chosen axis scales can create misleading apparent correlations or divergences

By adjusting the y-axis ranges independently, the designer can make any two series appear correlated or anti-correlated, misleading the audience about their true relationship.

Question 41: A data scientist is choosing between Random Forest and Gradient Boosting for a classification task on a large, noisy dataset. The project has a tight deadline, requiring a model that is relatively fast to train and less sensitive to hyperparameter tuning. Which algorithm is the better choice and why?

Gradient Boosting, because its sequential nature allows for faster convergence on the optimal solution.
Random Forest, because it only uses a single decision tree, making it computationally simple.
Random Forest, because its trees are built in parallel and it is generally less sensitive to hyperparameter changes. (Correct answer)
Gradient Boosting, because it can achieve higher accuracy, which is always the primary goal.

Correct answer: Random Forest, because its trees are built in parallel and it is generally less sensitive to hyperparameter changes.

Random Forest is the more suitable choice in this scenario. Its core design involves building many independent decision trees, a process that can be easily parallelized, leading to faster training times on multi-core systems. [10] It is also known to be more robust and less sensitive to hyperparameter tuning compared to Gradient Boosting, which often requires careful tuning to prevent overfitting. [6] While Gradient Boosting can sometimes achieve higher accuracy, the trade-off in training time and tuning complexity makes Random Forest better for this specific situation. [10]

Question 42: What is the primary purpose of a confidence interval for a regression coefficient?

To determine the R-squared value of the model
To estimate the range of plausible values for the true population parameter (Correct answer)
To test whether the model has multicollinearity
To predict individual future observations

Correct answer: To estimate the range of plausible values for the true population parameter

A confidence interval for a regression coefficient provides a range within which the true population parameter is likely to fall at a given confidence level.

Question 43: What is the 'elbow method' used to determine in unsupervised learning?

The number of principal components to retain in PCA
The optimal epsilon and minPts for DBSCAN
The convergence threshold for the EM algorithm
The appropriate number of clusters k in k-means by plotting inertia vs. k (Correct answer)

Correct answer: The appropriate number of clusters k in k-means by plotting inertia vs. k

The elbow method plots within-cluster sum of squares (inertia) against k; the 'elbow' point where improvement diminishes suggests the optimal k.

Question 44: A data scientist is preparing a presentation for a non-technical executive audience to explain the results of a complex customer segmentation model. Which of the following approaches is most effective for this type of communication?

Focus on the business impact and key takeaways, using simple, clear visuals and avoiding technical jargon. (Correct answer)
Provide the raw model output, including cluster centroids and silhouette scores, to ensure full transparency.
Explain the mathematical foundations of the clustering algorithm, starting with the distance metrics used.
Present a detailed walkthrough of the model's architecture, including the hyperparameter tuning process and feature engineering steps.

Correct answer: Focus on the business impact and key takeaways, using simple, clear visuals and avoiding technical jargon.

Effective communication requires tailoring the message to the audience. Executives are primarily concerned with business outcomes, ROI, and strategic implications, not the technical implementation details. A clear, concise presentation focused on the 'so what' is most effective.

Question 45: In generalized linear models (GLMs), the variance function describes:

The dispersion of predictor variables
How the variance of the response depends on its mean (Correct answer)
The penalty term applied to regression coefficients
The link between the linear predictor and the mean response

Correct answer: How the variance of the response depends on its mean

The variance function V(μ) in a GLM specifies how the variance of the response relates to its mean, distinguishing different members of the exponential family.

Question 46: In multiple regression, the partial F-test is used to:

Test each individual coefficient against zero
Test for heteroscedasticity in the residuals
Detect outliers in the response variable
Compare a reduced model to a full model by testing a subset of predictors simultaneously (Correct answer)

Correct answer: Compare a reduced model to a full model by testing a subset of predictors simultaneously

The partial F-test evaluates whether a group of predictors collectively adds significant explanatory power beyond what the reduced model already explains.

Question 47: What is the primary advantage of using UMAP over t-SNE for dimensionality reduction?

UMAP better preserves global structure and is significantly faster on large datasets (Correct answer)
UMAP requires labeled data to construct the embedding
UMAP uses a linear projection making it more interpretable
UMAP always produces two-dimensional outputs only

Correct answer: UMAP better preserves global structure and is significantly faster on large datasets

UMAP preserves more global structure than t-SNE and scales better computationally due to its graph-based approximation approach.

Question 48: Which of the following is a core limitation of PCA for dimensionality reduction in the context of unsupervised learning?

PCA only captures linear relationships and misses nonlinear structure in the data (Correct answer)
PCA requires specifying the number of components before seeing the data
PCA is computationally infeasible for datasets with fewer than 1,000 dimensions
PCA cannot handle datasets with more features than samples

Correct answer: PCA only captures linear relationships and misses nonlinear structure in the data

PCA finds orthogonal directions of maximum variance using linear projections, so it cannot capture curved manifolds or other nonlinear structure in high-dimensional data.

Question 49: What is the primary function of YARN in the Hadoop ecosystem?

SQL query processing
Distributed file storage
Cluster resource management and job scheduling (Correct answer)
Data serialization

Correct answer: Cluster resource management and job scheduling

YARN (Yet Another Resource Negotiator) manages cluster resources and schedules applications across the Hadoop cluster.

Question 50: Which technique is best suited for imputing missing values in a numerical column when the data is believed to be missing at random and the column has a moderate skew?

Median imputation (Correct answer)
Dropping all rows with missing values
Mean imputation
Forward fill

Correct answer: Median imputation

Median imputation is preferred over mean imputation when data is skewed because the median is robust to outliers and better represents the central tendency of a skewed distribution.

Question 51: Which storytelling framework structures a data presentation as: situation → complication → resolution?

McKinsey SCQA (Situation-Complication-Question-Answer) (Correct answer)
The Pyramid Principle
STAR method
Hero's Journey narrative arc

Correct answer: McKinsey SCQA (Situation-Complication-Question-Answer)

Barbara Minto's Pyramid Principle and McKinsey's SCQA both use Situation → Complication → Question → Answer to drive executive communication.

Question 52: Which probability theorem is foundational to generative classifiers like Naive Bayes, allowing posterior class probabilities to be computed from likelihoods and priors?

Jensen's Inequality
Central Limit Theorem
Bayes' Theorem (Correct answer)
Law of Large Numbers

Correct answer: Bayes' Theorem

Bayes' Theorem relates the posterior probability P(class|data) to the likelihood P(data|class) and the prior P(class). Naive Bayes classifiers apply this with the 'naive' assumption of conditional feature independence to make computation tractable.

Question 53: A data scientist is working with a dataset containing customer information, including an 'Income' feature with a significant number of missing values. The distribution of the 'Income' feature is heavily right-skewed. Which of the following methods for handling missing data is most appropriate in this situation to minimize the impact of outliers?

Deleting the rows with missing income values.
Replacing missing values with a constant like zero.
Imputing the missing values with the median of the 'Income' feature. (Correct answer)
Imputing the missing values with the mean of the 'Income' feature.

Correct answer: Imputing the missing values with the median of the 'Income' feature.

For skewed distributions, the median is a more robust measure of central tendency than the mean because it is less affected by outliers. Deleting rows could lead to significant data loss, and replacing with zero would introduce bias and distort the feature's distribution.

Question 54: In a convolutional neural network (CNN), what is the primary purpose of a pooling layer?

Normalize activations
Reduce spatial dimensions and computation (Correct answer)
Add non-linearity
Increase filter depth

Correct answer: Reduce spatial dimensions and computation

Pooling layers downsample feature maps, reducing spatial size and computational cost while providing a degree of translation invariance.

Question 55: A random sample of size n=36 is drawn from a population with σ=12. What is the standard error of the mean?

0.33
12
2 (Correct answer)
6

Correct answer: 2

Standard error = σ/√n = 12/√36 = 12/6 = 2.

Question 56: Which of the following languages should the question mark in the illustration below replace?

Java (Correct answer)
NLP
COBOL
PHP

Correct answer: Java

Without the specific illustration, it's impossible to provide context. However, Java is a widely used, high-level, class-based, object-oriented programming language known for its 'write once, run anywhere' capability. It is extensively used in enterprise-level applications, Android development, and big data technologies, making it a common language in data science contexts.

Question 57: What happens to the probability of a Type I error if a researcher lowers the significance level from 0.05 to 0.01?

It increases
It becomes equal to the probability of a Type II error
It decreases (Correct answer)
It remains the same

Correct answer: It decreases

The significance level α directly defines the probability of a Type I error (rejecting a true null hypothesis). Lowering α from 0.05 to 0.01 directly reduces the Type I error rate, though it simultaneously increases the risk of a Type II error.

Question 58: You train a logistic regression model and observe AUC-ROC = 0.97 on the training set but AUC-ROC = 0.61 on the test set. Which action most directly addresses this problem?

Increase the learning rate
Apply stronger regularization (increase λ) (Correct answer)
Use a larger batch size during training
Switch the evaluation metric to F1-score

Correct answer: Apply stronger regularization (increase λ)

A large gap between training and test AUC-ROC indicates overfitting (high variance). Increasing the regularization strength penalizes model complexity, reducing the gap between train and test performance.

Question 59: The most comprehensible type of text data is seen on social media platforms. A corpus of entire tweet social media data is provided to you. How can a model that suggests hashtags be made?

Train a word2vector model to learn repeating contexts in the sentences
Perform Topic Models to obtain most significant words of the corpus
Train a Bag of Ngrams model to capture top n-grams – words and their combinations
All of these (Correct answer)

Correct answer: All of these

To build a model that suggests hashtags from a corpus of tweet data, all the listed methods are valuable. Training a Word2Vec model helps learn semantic relationships and repeating contexts, enabling the suggestion of contextually relevant hashtags. Performing Topic Models identifies underlying themes and their most significant words, which can directly serve as hashtag candidates. Additionally, training a Bag of Ngrams model captures frequently occurring words and phrases, many of which are commonly used as hashtags.

Question 60: Which linkage criterion in hierarchical clustering tends to produce the most compact, spherical clusters?

Ward's method (Correct answer)
Complete linkage
Single linkage
Average linkage

Correct answer: Ward's method

Ward's method minimizes the total within-cluster variance at each merge step, producing compact, roughly spherical clusters.

Question 61: In spectral clustering, what is the primary role of the graph Laplacian matrix?

It captures the connectivity structure of the data for eigen-decomposition (Correct answer)
It performs dimensionality reduction via SVD
It normalizes the input features to unit variance
It computes the distance between all pairs of centroids

Correct answer: It captures the connectivity structure of the data for eigen-decomposition

The graph Laplacian encodes the similarity structure of the data as a graph, and its eigenvectors reveal the cluster structure used for partitioning.

Master of Data Science Exam

The Master of Data Science Exam exam validates essential knowledge and skills required for certification or licensure in this field.

Questions

60 min

Time Limit

70%

To Pass

Checking for saved progress...

Exam Rules

You can skip questions and return to them later
Flag questions for review before submitting
No feedback shown until you submit the entire exam
Unanswered questions count as wrong — answer everything
10 pretest questions are mixed in and don't affect your score
Timer auto-submits when time runs out
Your progress is auto-saved every 30 seconds