Chapter 22: Quiz
How Recommendation Algorithms Work: A Technical Introduction
Instructions: Select the best answer for each question. Answer key appears at the end.
1. Which of the following best describes the core difference between a rule-based recommendation system and a machine learning-based recommendation system?
A) Rule-based systems use more data than machine learning systems B) Rule-based systems have engineers specify what to recommend; machine learning systems learn from data what recommendations achieve a specified objective C) Machine learning systems are faster to build than rule-based systems D) Rule-based systems can only work for text content, not video
2. Content-based filtering recommends items by:
A) Finding users similar to the current user and recommending what those users liked B) Analyzing the features of items and recommending items similar to what the user has previously engaged with C) Decomposing the user-item interaction matrix into latent factors D) Randomly sampling items from the catalog and measuring engagement
3. Cosine similarity in the context of recommendation systems measures:
A) The number of shared interactions between two users B) The geographic distance between a user's location and content creator's location C) The angle between two feature vectors in high-dimensional space, indicating how similar two items are D) The time elapsed between a user's first and most recent platform interactions
4. The primary limitation of content-based filtering is that it:
A) Requires data from millions of users to function B) Cannot recommend items outside the feature space of what the user has already engaged with C) Is computationally too expensive to run in real time D) Cannot handle video content, only text and images
5. In collaborative filtering, the user-item interaction matrix is described as "sparse" because:
A) It only contains data from users in specific geographic regions B) Each user has interacted with only a tiny fraction of available items, leaving most matrix cells empty C) The matrix uses a compressed storage format to save computational memory D) Only verified users' data is included in the matrix
6. In Singular Value Decomposition (SVD) for recommendation systems, "latent factors" are:
A) Explicit content categories manually labeled by human moderators B) The hidden user identities that the algorithm discovers through behavioral patterns C) Unobserved characteristics automatically discovered from interaction patterns that explain preference structure D) Demographic variables like age and gender
7. The two-stage architecture used in industrial recommendation systems consists of:
A) Data collection and model training B) Candidate retrieval (narrow the catalog to thousands) followed by ranking (score each candidate with a richer model) C) User profiling and content profiling D) Exploitation and exploration phases
8. Multi-task learning in recommendation systems refers to:
A) Having multiple engineering teams work on the recommendation system simultaneously B) Training the model to simultaneously predict multiple engagement outcomes and combining them with weights C) Recommending content from multiple different content categories at once D) Running recommendation models on multiple device types
9. The "cold start problem" in recommendation systems refers to:
A) The slowdown in recommendation quality during winter months when user behavior changes B) The difficulty of making good recommendations for new users or new items that lack behavioral history C) The computational cost of starting a recommendation model from random initialization D) The challenge of recommending content to users who have disabled their usage history
10. YouTube's 2016 shift from optimizing for click-through rate to optimizing for watch time was intended to:
A) Reduce the computational cost of the recommendation system B) Reward shorter, higher-quality videos over longer, lower-quality videos C) Better capture genuine user interest by measuring how long users actually watched, not just whether they clicked D) Increase advertising revenue by showing more ads during longer videos
11. "Goodhart's Law" as applied to recommendation systems states that:
A) Algorithms become more accurate over time as they accumulate more data B) When a behavioral measure becomes a training objective, optimizing for it tends to diverge from optimizing for the underlying value it was intended to measure C) Good recommendation systems require at least 10 million users before they can personalize effectively D) Engagement metrics are always positively correlated with user wellbeing
12. The epsilon-greedy exploration strategy works by:
A) Gradually reducing the learning rate of the recommendation model over time B) With probability epsilon, recommending a random item for exploration; otherwise recommending the highest-scoring item C) Recommending items that score exactly at the epsilon percentile of predicted engagement D) Using user-specified preferences to set an exploration budget
13. Thompson Sampling is considered superior to epsilon-greedy for recommendation exploration because:
A) It is computationally faster B) It focuses exploration on items where the algorithm is most uncertain, rather than exploring randomly C) It requires less data to initialize D) It eliminates the need for any exploitation
14. "Preference amplification" as a feedback loop pathology means:
A) The algorithm artificially inflates engagement metrics to justify its recommendations B) Recommendations narrow toward increasingly specific content niches over time as the algorithm reinforces established engagement patterns C) Users' stated preferences amplify over time even as their behavioral preferences stay constant D) Content creators amplify their content production after receiving algorithmic boosts
15. The "distribution shift" problem in recommendation systems refers to:
A) The geographic shift in user base that occurs as platforms expand internationally B) As the algorithm changes its recommendations, it also changes the behavioral data it collects, making the system a moving target C) The gradual shift in content formats (from text to video) that changes how recommendations are computed D) The redistribution of recommendation weight from popular to niche content
16. Which of the following is NOT described in the chapter as a feature category used by recommendation algorithms?
A) User interaction history B) Item content features C) User biometric data (heart rate, galvanic skin response) D) Contextual features (time of day, device type)
17. The chapter states that "negative emotional states tend to produce stronger behavioral engagement signals than positive emotional states." This implies that engagement-optimized systems will:
A) Systematically filter out emotionally negative content to improve user experience B) Systematically favor emotionally arousing (including negative) content because it generates higher measured engagement C) Achieve perfect alignment between engagement and wellbeing because users naturally avoid negative content D) Require special algorithms to detect and suppress emotionally negative content
18. Facebook's "Meaningful Social Interactions" (MSI) metric was introduced in 2018. According to the chapter, what was the primary unintended consequence?
A) Users spent less time on Facebook, reducing advertising revenue B) The metric increased distribution of divisive and outrage-generating content because such content produced more comments and replies C) Small creators benefited while large media companies lost reach D) Users began sharing more personal information with Facebook
19. "Popularity bias" in recommendation systems refers to:
A) The tendency of recommendation algorithms to favor content from popular geographic regions B) A rich-get-richer dynamic where popular content accumulates more engagement data, leading to more recommendations, leading to more popularity C) The preference of popular users for certain types of algorithmic recommendations D) The tendency of algorithms to recommend content from popular creators regardless of individual user preferences
20. The chapter identifies the "unmeasured objective" as a key source of misalignment between algorithmic optimization and user wellbeing. The "unmeasured objective" refers to:
A) Hidden engagement metrics that platforms use but do not disclose publicly B) User flourishing — whether content helps users understand the world, strengthens relationships, or improves wellbeing — which cannot be measured automatically at scale C) The revenue targets that platforms optimize for alongside engagement metrics D) The quality ratings provided by human content reviewers
21. Maya's late-night TikTok session at 11:15 pm is described in the chapter. Which of the following information does the algorithm NOT have access to?
A) Her device type and location B) Her completion rates for videos in the current session C) Whether she has homework due the next morning D) The recency-weighted pattern of her recent likes
22. The chapter argues that the misalignment between engagement optimization and wellbeing optimization is "structural rather than incidental." This means:
A) The misalignment results from intentional decisions by platform engineers to prioritize revenue over wellbeing B) The misalignment arises from the architecture of the systems themselves — the choice of what to measure, what to optimize, and how training data is generated — rather than from individual bad decisions C) The misalignment can be fixed by adding better content moderation policies D) The misalignment is caused by users who actively choose to engage with harmful content
Answer Key
- B
- B
- C
- B
- B
- C
- B
- B
- B
- C
- B
- B
- B
- B
- B
- C
- B
- B
- B
- B
- C
- B