Chapter 10 Quiz: Recommendation Systems

Multiple Choice

Question 1. Amazon estimates that approximately what percentage of its revenue is generated through its recommendation engine?

a) 10 percent b) 20 percent c) 35 percent d) 50 percent

Question 2. In collaborative filtering, the user-item interaction matrix is typically characterized by which property?

a) It is dense — most cells contain ratings b) It is symmetric — user similarity equals item similarity c) It is extremely sparse — less than 1 percent of cells are filled in large systems d) It is binary — cells contain only 0 or 1

Question 3. What is the primary advantage of item-based collaborative filtering over user-based collaborative filtering?

a) Item-based filtering requires less data b) Item similarities are more stable over time than user similarities c) Item-based filtering can handle the cold start problem for new items d) Item-based filtering produces more diverse recommendations

Question 4. A user rates movies on a 1-5 scale with an average rating of 4.5 (a "generous rater"). Another user rates movies on the same scale with an average of 2.1 (a "tough grader"). Both users rank the same movies in the same order. Which similarity metric would correctly identify them as having identical preferences?

a) Cosine similarity b) Pearson correlation c) Jaccard similarity d) Euclidean distance

Question 5. In matrix factorization for recommendation systems, what are "latent factors"?

a) Observable product features such as price, brand, and category b) Hidden dimensions discovered by the algorithm that explain patterns in user-item interactions c) The number of users and items in the ratings matrix d) Missing values in the user-item interaction matrix

Question 6. Which of the following best describes the key lesson from the Netflix Prize competition?

a) Deep learning always outperforms traditional methods for recommendation b) A single well-tuned algorithm is better than an ensemble c) Matrix factorization was a core technique, but Netflix never fully deployed the winning solution due to engineering complexity d) The winning team used pure content-based filtering

Question 7. A content-based recommendation system for an e-commerce site has been deployed for six months. Users complain that they keep seeing "the same types of products." Which limitation of content-based filtering does this illustrate?

a) The cold start problem b) Overspecialization (filter bubble) c) Data sparsity d) Scalability constraints

Question 8. In a switching hybrid recommendation system, which approach is typically used for new users with no interaction history?

a) Collaborative filtering b) Matrix factorization c) Popularity-based or content-based filtering d) Bayesian personalized ranking

Question 9. The cold start problem refers to which challenge?

a) The recommendation system takes too long to compute results b) The system lacks sufficient data to make quality recommendations for new users or new items c) User preferences change over time, making past data unreliable d) The system recommends items that are no longer available

Question 10. Which of the following is an example of implicit feedback?

a) A user rates a product 4 out of 5 stars b) A user writes a product review c) A user spends 3 minutes viewing a product page without purchasing d) A user marks an item as a "favorite"

Question 11. Why is the absence of implicit feedback (e.g., a user not clicking on an item) NOT a reliable negative signal?

a) Users always click on items they see b) The user may not have seen the item, the thumbnail may have been unappealing, or the user may have been distracted c) Implicit feedback is always positive d) Negative signals are more reliable than positive ones

Question 12. A recommendation system returns 10 items to a user. The user interacts with 4 of them. Of the 20 items the user eventually interacted with during the session, 4 were in the recommendation list. What is the Precision@10 and Recall@10?

a) Precision@10 = 0.40, Recall@10 = 0.40 b) Precision@10 = 0.40, Recall@10 = 0.20 c) Precision@10 = 0.20, Recall@10 = 0.40 d) Precision@10 = 0.20, Recall@10 = 0.20

Question 13. NDCG (Normalized Discounted Cumulative Gain) differs from Precision@K in that NDCG:

a) Only considers the top 5 recommendations b) Gives more credit when relevant items appear higher in the ranking c) Penalizes systems that recommend popular items d) Measures diversity rather than accuracy

Question 14. A recommendation system achieves Precision@10 of 0.45 but a catalog coverage of only 3 percent. What is the most likely problem?

a) The system is too diverse b) The system is recommending mostly popular items (popularity bias) c) The system has a cold start problem d) The system is using content-based filtering

Question 15. Which of the following is the most accurate description of a "filter bubble" in the context of recommendation systems?

a) A spam filter that blocks legitimate recommendations b) A narrowing of recommendations over time as the system reinforces existing preferences and hides alternatives c) A technical failure where the system returns empty recommendations d) A privacy mechanism that anonymizes user data

Question 16. In Athena's hybrid recommendation architecture, what is the role of the nightly batch pipeline versus the real-time component?

a) The batch pipeline handles new users; the real-time component handles returning users b) The batch pipeline generates a candidate pool for each customer; the real-time component re-ranks candidates based on the current session c) The batch pipeline computes content-based recommendations; the real-time component computes collaborative filtering d) The batch pipeline trains the model; the real-time component retrains it on new data

Question 17. NK noticed that Athena's recommendation engine disproportionately surfaced premium products to customers in certain zip codes. This occurred because:

a) The merchandising team manually biased the recommendations b) Collaborative filtering learned from historical purchasing patterns that reflected socioeconomic differences c) The content-based model was trained on biased product descriptions d) The recommendation system was deliberately programmed with zip-code-based pricing rules

Question 18. Which recommendation evaluation metric specifically captures the ability of a system to surface items that are both surprising AND relevant to the user?

a) Precision@K b) NDCG c) Coverage d) Serendipity

Question 19. A grocery delivery app uses batch recommendations updated every 6 hours. Which scenario would be MOST poorly served by this approach?

a) A customer who orders the same items every week b) A customer searching for ingredients for a recipe they just found c) A customer who always shops on Sunday mornings d) A customer who prefers organic products

Question 20. In the RecommendationEngine class, what is the purpose of the _popular_items method?

a) To compute item-item similarity scores b) To serve as a fallback for cold-start users who have no interaction history c) To evaluate the precision of the recommendation engine d) To filter out unpopular items from the candidate pool

Short Answer

Question 21. Tom Kowalski explains latent factors to a non-technical executive by comparing them to "hidden DNA." In two to three sentences, explain why this metaphor is effective and identify one way in which it is misleading.

Question 22. Professor Okonkwo states that "recommendation systems are the difference between a catalog and a conversation." In three to four sentences, explain what she means by this distinction and why it matters strategically for Athena Retail Group.

Question 23. Explain the tension between optimizing for accuracy (Precision@K) and optimizing for diversity/novelty in a recommendation system. Use a concrete e-commerce example to illustrate why a company might deliberately sacrifice some accuracy to improve diversity.

Question 24. The chapter describes Athena's "algorithmic curation with editorial override" compromise. In two to three sentences, explain why this approach is strategically superior to either pure algorithmic recommendations or pure editorial curation.

Question 25. NK asks: "Are we recommending what customers want, or are we manipulating what they want?" In three to four sentences, explain why this question is difficult to answer and identify one specific design decision that would move a recommendation system from "serving" toward "manipulating."

Answer key is available in Appendix B (Answers to Selected Exercises).