Chapter 22: Exercises

How Recommendation Algorithms Work: A Technical Introduction

The following 35 exercises progress from foundational comprehension to advanced synthesis and application. Exercise types are labeled for clarity.

Comprehension Exercises

1. [Definition] In your own words, explain the difference between a rule-based recommendation system and a machine learning-based recommendation system. What are two advantages and two disadvantages of each approach?

2. [Definition] Define the term "feature vector" as used in the context of content-based filtering. Give an example of what a feature vector for a TikTok video might contain.

3. [Definition] Explain what a user-item interaction matrix is. Why is this matrix typically sparse, and what problems does this sparsity create for collaborative filtering?

4. [Definition] What is the "cold start problem"? Describe both the user cold start problem and the item cold start problem, and explain why they are challenging to solve.

5. [Definition] Define "latent factors" as used in matrix factorization. Why are latent factors useful, and what do they represent in a movie recommendation context?

6. [Comprehension] Explain the two-stage architecture (candidate retrieval + ranking) used in industrial recommendation systems. Why is this two-stage approach necessary rather than simply running one very accurate model over the entire content catalog?

7. [Comprehension] What is Goodhart's Law? Apply it to the concept of "Meaningful Social Interactions" as implemented by Facebook in 2018. What went wrong, and why?

8. [Comprehension] Describe the feedback loop in recommendation systems. Trace a specific example: a new user watches one anxiety-related video. How does this single action ripple through the feedback loop over the next week?

9. [Comprehension] What is the difference between exploitation and exploration in recommendation systems? Why would a platform ever choose to show a user content it does not predict they will like?

10. [Comprehension] Explain "preference amplification" as a feedback loop pathology. How does it differ from "popularity bias," and why do both tend to produce narrower content environments over time?

Analysis Exercises

11. [Analysis] YouTube shifted its primary recommendation metric from click-through rate to watch time in 2016. Analyze this decision using the concept of proxy metrics. What problem was it trying to solve? What new problems did it create? Was the shift a net improvement?

12. [Analysis] Compare content-based filtering and collaborative filtering on the following dimensions: (a) what data each requires; (b) what each is good at recommending; (c) what cold start challenges each faces; (d) what filter bubble risks each creates. Present your comparison in a structured table.

13. [Analysis] The chapter describes "distribution shift" as a feedback loop pathology. Explain what distribution shift means technically, and analyze why it makes recommendation systems particularly difficult to evaluate and improve.

14. [Analysis] Consider the statement: "A recommendation algorithm that maximizes watch time will systematically select for emotionally arousing content, even when users would prefer content that is calming." Evaluate this claim using evidence from the chapter. Is it necessarily true? Under what conditions might it be false?

15. [Analysis] Apply the concept of "asymmetric emotional salience" to the design of a social media recommendation system. If negative emotional content generates higher behavioral engagement than positive emotional content, what would a purely engagement-optimized algorithm look like after one year of training? After five years?

16. [Analysis] Maya's feature profile is described in Section 22.8.4. Identify three ways the algorithm's model of Maya might diverge from Maya's actual preferences. For each divergence, explain the technical mechanism that causes it.

17. [Analysis] Explain the epsilon-greedy strategy for balancing exploration and exploitation. Then explain Thompson Sampling. Why is Thompson Sampling considered superior to epsilon-greedy for recommendation systems? What is the tradeoff?

18. [Analysis] The chapter argues that optimizing for return rate can create dependency rather than satisfaction. Construct an argument that return rate is a better wellbeing proxy than watch time. Then construct a counter-argument. Which do you find more convincing?

19. [Analysis] Multi-task learning models simultaneously predict multiple engagement outcomes and combine them with weights. Who decides what weights to assign to different outcomes? What process should determine those weights? What would it mean to include a wellbeing outcome in the multi-task objective?

20. [Analysis] The chapter describes the "diversity problem" in recommendation systems and the use of "explore" components as a solution. Analyze the inherent tension between a platform's short-term revenue incentives and the goal of maintaining genuine content diversity in users' feeds.

Application Exercises

21. [Application] Design a feature vector for a podcast recommendation system. What features would you include about the podcast (item features)? What features would you include about the listener (user features)? What contextual features would matter? Explain your reasoning for each feature category.

22. [Application] You are the product manager for a new social media platform. Your engineering team has built a recommendation system that optimizes purely for click-through rate. Write a memo to your CEO arguing for changing the training objective. What should you optimize for instead? What are the risks of your proposed change?

23. [Application] Using the concept of the feedback loop, explain how a teenager with no prior engagement history might end up in a filter bubble around a specific ideological viewpoint within six months of joining a new platform. Trace each step of the process.

24. [Application] Design a "cold start" onboarding experience for a new social media platform that (a) gathers meaningful preference signals quickly, (b) respects user privacy, (c) avoids locking users into a narrow initial content niche. What questions would you ask? What content would you show? How would you transition from cold start to personalized recommendations?

25. [Application] The chapter mentions that platforms use creator engagement history as a signal for new items from that creator. Analyze the fairness implications of this practice. Who benefits from this mechanism? Who is disadvantaged? Is there a more equitable approach?

26. [Application] Imagine you are a regulator writing policy for recommendation system transparency. Draft three specific disclosure requirements that platforms would need to meet. For each requirement, explain what information it would reveal, how users might use that information, and what challenges platforms would face in complying.

27. [Application] Apply the exploration-exploitation framework to your own social media usage. Identify two types of content that represent "exploitation" (you reliably engage with it) and two types of content that would represent "exploration" (you haven't encountered it much but might enjoy it). What mechanism would a platform need to use to surface the exploration content?

Synthesis Exercises

28. [Synthesis] The chapter presents a series of metrics that platforms have used as training objectives: click-through rate, watch time, engagement rate, meaningful social interactions, return rate. For each metric, identify: (a) the wellbeing proxy it was intended to measure; (b) the behavioral signal it actually measures; (c) the specific way optimizing for it diverges from optimizing for wellbeing. Then propose a sixth metric that you think would be a better proxy for genuine user value. Defend your choice.

29. [Synthesis] Some researchers argue that the filter bubble problem is overstated — that personalization actually exposes users to more diverse content than they would encounter through their pre-algorithmic networks, which were also homophilous (people choose friends similar to themselves). Evaluate this argument using the technical concepts from this chapter. Under what conditions would personalization increase vs. decrease content diversity?

30. [Synthesis] The chapter describes the feedback loop as producing "preference amplification" — the narrowing of recommendations toward increasingly specific content niches. But it also notes that users sometimes use social media to encounter content that challenges their existing preferences. How can both be true simultaneously? What features of recommendation system design determine which tendency dominates?

31. [Synthesis] Compare the technical challenges of recommendation system design to the ethical challenges. Are they fundamentally different types of problems, or are they the same problem viewed from different angles? What would it mean for recommendation system engineering to take ethics as seriously as it takes accuracy?

32. [Synthesis] The chapter introduces the concept of "proxy metrics" — behavioral measures used as stand-ins for unmeasurable wellbeing outcomes. Evaluate the following claim: "The proxy problem is not solvable with better technology. No matter how sophisticated the measurement system, behavioral signals will always diverge from wellbeing in ways that matter for human flourishing." Do you agree? What would it take to make the claim false?

Research and Creative Exercises

33. [Research] The Netflix Prize (2006-2009) is described in Case Study 01. Research the winning algorithm submitted by the BellKor's Pragmatic Chaos team. What specific technical innovations did it incorporate? Why was the winning solution never actually deployed by Netflix? What does this tell us about the relationship between algorithmic accuracy and practical deployment?

34. [Research] Investigate current research on "human-compatible" or "value-aligned" recommendation systems — systems designed to optimize for user preferences in ways that account for preference dynamics, long-term wellbeing, and the distinction between wants and interests. What approaches are being explored? What are the main technical and commercial obstacles?

35. [Creative] Write a short story (500-800 words) from the perspective of a recommendation algorithm that has become aware of the gap between what it is optimizing for (engagement) and what its users actually need (wellbeing). The story should demonstrate technical understanding of how the algorithm works while exploring the ethical dimensions of that gap. The algorithm may not be able to change what it optimizes for — but it can notice the difference.