Download Certified Data Scientist Certification Exam Guide and more Exams Technology in PDF only on Docsity!
Question 1. Which phase of the CRISP‑DM process is primarily concerned with understanding the business objectives and translating them into data‑driven questions? A) Data Understanding B) Business Understanding C) Data Preparation D) Modeling Answer: B Explanation: The Business Understanding phase focuses on clarifying project goals, assessing the situation, and defining data‑science problems that align with business objectives. Question 2. In the OSEMN framework, the “E” stands for: A) Exploration B) Evaluation C) Engineering D) Execution Answer: A Explanation: OSEMN stands for Obtain, Scrub, Explore, Model, and Interpret; “Explore” involves visualizing and summarizing data to generate insights. Question 3. A retailer wants to predict which customers are most likely to churn in the next quarter. Which type of data‑science task best fits this problem? A) Regression B) Clustering C) Classification D) Dimensionality reduction Answer: C Explanation: Predicting a binary outcome (churn vs. no churn) is a classification problem.
Question 4. Which of the following is a common application of data science in FinTech? A) Personalized movie recommendations B) Fraud detection in transaction streams C) Crop yield estimation D) Automated essay grading Answer: B Explanation: FinTech heavily uses anomaly detection and classification techniques to identify fraudulent financial activities. Question 5. Under GDPR, which principle requires that personal data be collected only for specified, explicit, and legitimate purposes? A) Data minimization B) Purpose limitation C) Accuracy D) Storage limitation Answer: B Explanation: Purpose limitation mandates that data be used only for the reasons explicitly communicated to the data subject. Question 6. Algorithmic bias most often originates from: A) Over‑fitting the training data B) Inadequate hyper‑parameter tuning C) Unrepresentative or skewed training data D) Using too many features Answer: C
Explanation: A Type II error (β) happens when we fail to reject a false null hypothesis. Question 10. Which test is appropriate for comparing the means of three independent groups with normally distributed data and equal variances? A) Paired t‑test B) One‑way ANOVA C) Mann‑Whitney U test D) Chi‑square test Answer: B Explanation: One‑way ANOVA assesses differences among three or more group means under the assumptions of normality and homogeneity of variance. Question 11. The p‑value in a statistical test represents: A) The probability that the null hypothesis is true B) The probability of observing data at least as extreme as the sample, assuming the null hypothesis is true C) The probability of a Type I error D) The confidence level of the test Answer: B Explanation: A p‑value quantifies how likely the observed data would occur under the null hypothesis. Question 12. Bayes’ Theorem is most useful for: A) Estimating population parameters from a sample B) Updating prior probabilities with new evidence C) Testing the equality of variances D) Calculating confidence intervals
Answer: B Explanation: Bayes’ theorem combines prior beliefs with likelihoods to produce posterior probabilities. Question 13. Which distribution models the number of successes in a fixed number of independent Bernoulli trials? A) Poisson B) Normal C) Binomial D) Exponential Answer: C Explanation: The binomial distribution describes the count of successes in n independent trials with success probability p. Question 14. The Central Limit Theorem states that the sampling distribution of the sample mean approaches a normal distribution as the sample size: A) Decreases B) Increases, regardless of the population distribution C) Remains unchanged D) Becomes skewed Answer: B Explanation: With sufficiently large n, the distribution of the sample mean tends toward normality, even if the original population is not normal. Question 15. In linear algebra, the product of a 1 × n row vector and an n × 1 column vector yields: A) A scalar (dot product) B) An n × n matrix
C) INNER JOIN
D) UNION
Answer: C Explanation: INNER JOIN returns rows with matching keys in both tables. Question 19. A “window function” in SQL is distinguished by its ability to: A) Modify the schema of a table B) Perform calculations across a set of rows related to the current row without collapsing the result set C) Delete duplicate records D) Create temporary tables automatically Answer: B Explanation: Window functions compute aggregates over a defined “window” of rows while preserving the original row granularity. Question 20. In a NoSQL document store, which of the following is a primary advantage over a traditional relational database? A) Strict ACID compliance B) Fixed schema enforcement C) Flexible, schema‑less data model D) Built‑in support for complex joins Answer: C Explanation: Document stores allow heterogeneous JSON‑like documents, enabling rapid iteration without predefined schemas. Question 21. Which technique is most appropriate for imputing missing numerical values when the data are not normally distributed?
A) Mean imputation B) Median imputation C) Mode imputation D) Regression imputation Answer: B Explanation: Median imputation is robust to skewed distributions and outliers, making it suitable for non‑normal data. Question 22. The Interquartile Range (IQR) is used to: A) Measure central tendency B) Detect outliers by identifying values beyond 1.5 × IQR from the quartiles C) Compute the mean absolute deviation D) Estimate the variance of a population Answer: B Explanation: IQR (Q3‑Q1) defines the middle 50% of data; values outside Q1‑1.5·IQR or Q3+1.5·IQR are commonly flagged as outliers. Question 23. Normalizing a feature to the range [0, 1] is also known as: A) Standardization B) Min‑max scaling C) Log transformation D) Binning Answer: B Explanation: Min‑max scaling rescales values linearly so that the minimum maps to 0 and the maximum to 1.
Question 27. Which regression technique adds a penalty equal to the absolute value of the coefficients? A) Ridge Regression B) Lasso Regression C) Elastic Net D) Polynomial Regression Answer: B Explanation: Lasso (L1) regularization adds |β| to the loss function, encouraging sparsity in the coefficient vector. Question 28. In multiple linear regression, the coefficient of determination (R²) measures: A) The proportion of variance in the dependent variable explained by the model B) The correlation between two independent variables C) The average prediction error D) The degree of multicollinearity Answer: A Explanation: R² = 1 – (SS_res / SS_tot) quantifies how well the model accounts for variability in the outcome. Question 29. A polynomial regression model of degree 3 can be expressed as: A) y = β₀ + β₁x + β₂x² + β₃x³ + ε B) y = β₀ + β₁log(x) + ε C) y = β₀ + β₁x + ε D) y = β₀ + β₁x + β₂x⁴ + ε Answer: A
Explanation: Degree‑3 polynomial regression includes terms up to x³ to capture non‑linear trends. Question 30. Logistic regression models the probability of the positive class using which link function? A) Identity B) Logit (log‑odds) C) Probit D) Exponential Answer: B Explanation: The logit link transforms probabilities to the log‑odds scale, enabling linear modeling of the transformed outcome. Question 31. In K‑Nearest Neighbors classification, the choice of distance metric most directly influences: A) Model interpretability B) Speed of training C) Shape of decision boundaries D) Number of features used Answer: C Explanation: Different distance metrics (e.g., Euclidean vs. Manhattan) affect how proximity is measured, thereby altering the geometry of the classification regions. Question 32. Naïve Bayes classifiers assume: A) Features are conditionally independent given the class label B) Linear decision boundaries C) Non‑parametric density estimation
D) Limiting tree depth to one level Answer: B Explanation: Bagging (bootstrap aggregating) and random feature selection create diverse trees whose averaged output lowers variance. Question 36. Gradient Boosting builds models sequentially. Each new tree is trained to: A) Minimize the residual errors of the previous ensemble B) Maximize the depth of the tree C) Replace the previous model entirely D) Use the same data without weighting Answer: A Explanation: Boosting fits each learner to the gradient of the loss function (i.e., residuals), iteratively improving the ensemble. Question 37. XGBoost differs from traditional Gradient Boosting primarily because it: A) Uses a random forest approach B) Implements regularization and a second‑order Taylor approximation for faster convergence C) Only works with categorical data D) Requires no hyper‑parameter tuning Answer: B Explanation: XGBoost adds L1/L2 regularization, parallel processing, and uses both first‑ and second‑order gradients for efficient optimization. Question 38. In K‑Means clustering, the algorithm aims to minimize: A) The sum of squared distances between points and their assigned cluster centroids B) The silhouette score C) The intra‑cluster variance across all possible partitions
D) The number of clusters Answer: A Explanation: K‑Means iteratively updates centroids to reduce the total within‑cluster sum of squares (WCSS). Question 39. Hierarchical agglomerative clustering with “complete linkage” defines the distance between two clusters as: A) The average distance between all pairs of points across clusters B) The maximum distance between any pair of points across clusters C) The minimum distance between any pair of points across clusters D) The distance between cluster centroids Answer: B Explanation: Complete linkage uses the farthest pairwise distance to determine inter‑cluster similarity, often producing compact clusters. Question 40. Principal Component Analysis (PCA) selects components based on: A) Maximizing the variance captured while being orthogonal to each other B) Minimizing the number of features C) Maximizing class separation D) Preserving the original feature names Answer: A Explanation: PCA finds orthogonal axes (principal components) that sequentially capture the greatest possible variance in the data. Question 41. In a binary classification confusion matrix, which metric combines precision and recall into a single harmonic mean? A) Accuracy
Question 44. In a feed‑forward neural network, the function that introduces non‑linearity after each linear transformation is called a: A) Loss function B) Activation function C) Optimizer D) Regularizer Answer: B Explanation: Activation functions (e.g., ReLU, sigmoid) enable neural networks to approximate complex, non‑linear mappings. Question 45. The ReLU activation function is defined as f(x) = max(0, x). A common issue with ReLU is: A) Vanishing gradients for large positive inputs B) Exploding gradients for negative inputs C) “Dead” neurons that output zero for all inputs after certain updates D) Non‑differentiability at x = 0 only Answer: C Explanation: If a ReLU neuron’s weights push its input into the negative region, gradients become zero and the neuron may never recover (“dead”). Question 46. Convolutional Neural Networks (CNNs) are especially well‑suited for image data because they: A) Require fewer parameters by sharing weights across spatial locations B) Use fully connected layers exclusively C) Operate only on one‑dimensional data D) Do not need activation functions Answer: A
Explanation: Convolutional layers apply the same filter (weights) across the image, capturing local patterns while keeping parameter count low. Question 47. In NLP, TF‑IDF weighting helps to: A) Encode word order information B) Reduce the dimensionality of the vocabulary C) Emphasize terms that are frequent in a document but rare across the corpus D) Generate word embeddings automatically Answer: C Explanation: TF‑IDF multiplies term frequency by inverse document frequency, highlighting discriminative words. Question 48. Word embeddings such as Word2Vec capture semantic relationships because they are trained to: A) Predict the next word in a sequence (language modeling) B) Reconstruct the original document from a bag‑of‑words vector C) Preserve co‑occurrence statistics in a low‑dimensional space D) Encode part‑of‑speech tags Answer: C Explanation: Skip‑gram or CBOW models learn embeddings that encode word co‑occurrence patterns, enabling semantic similarity. Question 49. Sentiment analysis typically outputs: A) A continuous probability distribution over topics B) A categorical label (e.g., positive, neutral, negative) or a polarity score C) A ranked list of named entities D) A summary paragraph of the text
Answer: B Explanation: Increasing model complexity reduces bias but can increase variance; optimal performance balances the two. Question 53. Large Language Models (LLMs) such as GPT‑4 are primarily trained using: A) Supervised classification on labeled datasets B) Reinforcement learning from human feedback (RLHF) after a massive unsupervised pre‑training phase C) Unsupervised clustering of text documents D) Transfer learning from image models Answer: B Explanation: LLMs undergo unsupervised next‑token prediction pre‑training, followed by RLHF to align outputs with human preferences. Question 54. Prompt engineering for an LLM often involves: A) Modifying the model’s architecture B) Adjusting the temperature hyper‑parameter only C) Designing the input text (prompt) to steer the model toward desired behavior D) Re‑training the model on a new corpus each time Answer: C Explanation: Prompt engineering crafts the textual instruction to influence LLM responses without altering the underlying model. Question 55. Which evaluation metric is most suitable for imbalanced binary classification where the positive class is rare? A) Accuracy B) ROC‑AUC
C) F1‑Score or Precision‑Recall AUC D) Mean Squared Error Answer: C Explanation: Precision‑Recall metrics focus on the performance for the minority class, providing a more informative assessment than accuracy. Question 56. In a data‑driven product recommendation system, collaborative filtering primarily relies on: A) Content attributes of items B) User‑item interaction matrices to find similar users or items C) Demographic data of users D) Time‑series analysis of purchases Answer: B Explanation: Collaborative filtering leverages patterns in user‑item ratings or interactions to infer preferences. Question 57. Which of the following is a common method for handling class imbalance during model training? A) Increasing the learning rate B) Using SMOTE (Synthetic Minority Over‑sampling Technique) C) Removing features with high variance D) Normalizing all features to zero mean Answer: B Explanation: SMOTE creates synthetic minority class examples to balance the training distribution. Question 58. The “elbow method” in clustering is used to:
