MAC Predictive Modeling 3 — Questions and Answers
Question 1: Which regularization technique sets some regression coefficients exactly to zero, effectively performing feature selection?
- Ridge (L2)
- LASSO (L1) (Correct answer)
- Elastic Net
- Dropout
Correct answer: LASSO (L1)
LASSO regression uses an L1 penalty that drives some coefficients to exactly zero, automatically selecting the most relevant features.
Question 2: When building a propensity-to-buy model, the target variable is typically:
- Average order value
- Binary — purchased or did not purchase (Correct answer)
- Customer lifetime value score
- Time between purchases
Correct answer: Binary — purchased or did not purchase
Propensity-to-buy models predict the probability of a binary outcome: whether a customer will purchase (1) or not (0).
Question 3: What does an AUC-ROC score of 0.5 indicate about a classification model?
- The model achieves 50% accuracy on training data
- The model performs no better than random chance (Correct answer)
- The model perfectly separates positive and negative classes
- The model has 50% precision and 50% recall
Correct answer: The model performs no better than random chance
An AUC of 0.5 means the model's ROC curve follows the diagonal, indicating it performs no better than random guessing.
Question 4: In gradient boosting, how are subsequent models trained?
- Independently on random subsets of data
- On the residual errors of previous models (Correct answer)
- Using the average predictions of all prior models
- By increasing the weight of correctly classified samples
Correct answer: On the residual errors of previous models
Gradient boosting trains each new model to correct the residual errors (pseudo-residuals) left by the combined ensemble of previous models.
Question 5: A marketer wants to predict next month's email campaign revenue. Which type of model is most appropriate?
- Binary classification
- Multi-class classification
- Regression (Correct answer)
- Clustering
Correct answer: Regression
Predicting a continuous value like revenue requires a regression model, not a classification model.
Question 6: What is the key advantage of using cross-validation over a single train/test split?
- It trains the model faster by reducing dataset size
- It provides a more reliable performance estimate by using all data for both training and testing (Correct answer)
- It eliminates the need for hyperparameter tuning
- It prevents the model from using categorical variables
Correct answer: It provides a more reliable performance estimate by using all data for both training and testing
Cross-validation rotates the test set across multiple folds so every observation is used for testing, yielding a more reliable and less variance-prone performance estimate.
Question 7: Which of the following best describes 'feature engineering' in predictive modeling?
- Selecting the algorithm with the lowest error rate
- Creating or transforming input variables to improve model performance (Correct answer)
- Removing outliers from the dataset before training
- Tuning the learning rate of a neural network
Correct answer: Creating or transforming input variables to improve model performance
Feature engineering involves creating new variables, transforming existing ones, or combining features to give the model better predictive signals.
MAC Predictive Modeling 3
Which regularization technique sets some regression coefficients exactly to zero, effectively performing feature selection?