MS Master of Data Science 3 — Questions and Answers
Question 1: Which ensemble method trains multiple models sequentially, where each model focuses on correcting the errors of the previous one?
- Bagging
- Random Forest
- Boosting (Correct answer)
- Stacking
Correct answer: Boosting
Boosting trains models sequentially, with each subsequent model assigning higher weight to misclassified instances from previous models.
Question 2: In natural language processing, what does TF-IDF stand for and what problem does it solve?
- Term Frequency-Inverse Document Frequency; it downweights common words across documents (Correct answer)
- Text Filtering-Index Document Format; it speeds up text search
- Token Feature-Information Density Factor; it measures sentence complexity
- Total Frequency-Inverted Data Format; it normalizes text length
Correct answer: Term Frequency-Inverse Document Frequency; it downweights common words across documents
TF-IDF weights terms by how often they appear in a document (TF) relative to how rare they are across all documents (IDF), reducing the importance of common stop words.
Question 3: A model achieves 99% accuracy on a dataset where 99% of records belong to class A. What metric would better reveal poor performance on class B?
- Overall accuracy
- F1-score on the minority class (Correct answer)
- Mean squared error
- R-squared
Correct answer: F1-score on the minority class
The F1-score on the minority class captures both precision and recall for underrepresented classes, exposing poor performance that high overall accuracy masks.
Question 4: What is the curse of dimensionality and its main effect on distance-based algorithms like KNN?
- Too many features slow down training only
- In high dimensions, all points become approximately equidistant, making distance metrics meaningless (Correct answer)
- Having more dimensions always improves model accuracy
- High dimensionality only affects linear models
Correct answer: In high dimensions, all points become approximately equidistant, making distance metrics meaningless
As dimensions increase, distances between data points converge, so distance-based algorithms like KNN lose their discriminative power.
Question 5: Which data warehouse schema stores a central fact table surrounded by denormalized dimension tables, optimizing for query performance?
- Snowflake schema
- Star schema (Correct answer)
- Galaxy schema
- OLTP schema
Correct answer: Star schema
The star schema places a fact table at the center with denormalized dimension tables radiating outward, minimizing joins and speeding up analytical queries.
Question 6: In time series analysis, what does 'stationarity' mean and why is it important?
- The series has no trend; stationarity is required for most time series models like ARIMA (Correct answer)
- The series has constant values; stationarity simplifies data storage
- The series is collected at fixed intervals; stationarity enables forecasting
- The series has no noise; stationarity improves visualization
Correct answer: The series has no trend; stationarity is required for most time series models like ARIMA
A stationary series has constant mean, variance, and autocorrelation over time, which is assumed by classical models like ARIMA for valid forecasting.
Question 7: What is transfer learning in deep learning, and when is it most beneficial?
- Copying model weights between identical architectures; useful when models are too large
- Reusing a pre-trained model's learned features for a new but related task; most useful with limited labeled data (Correct answer)
- Moving trained models between cloud providers; useful for cost savings
- Training on synthetic data then testing on real data; useful for rare events
Correct answer: Reusing a pre-trained model's learned features for a new but related task; most useful with limited labeled data
Transfer learning applies features learned from large datasets (e.g., ImageNet) to new tasks, dramatically reducing the labeled data and compute needed.
MS Master of Data Science 3
Which ensemble method trains multiple models sequentially, where each model focuses on correcting the errors of the previous one?