Chapter 22: Further Reading

How Recommendation Algorithms Work: A Technical Introduction

The following annotated sources are organized by theme. Academic papers are noted where applicable; popular press sources are included where they provide unique primary information.

Foundational Technical Works

1. Koren, Y., Bell, R., & Volinsky, C. (2009). "Matrix Factorization Techniques for Recommender Systems." IEEE Computer, 42(8), 30-37. The definitive academic overview of matrix factorization for recommendation systems, written by members of the BellKor team that won the Netflix Prize. Accessible to readers without advanced mathematics training while covering the core technical concepts. The most-cited paper in recommendation systems research and an essential reference for anyone who wants to understand why latent factor models outperform neighborhood methods.

2. Funk, Simon (2006). "Netflix Update: Try This at Home." Simon Funk's Blog. The blog post that seeded the Netflix Prize field with matrix factorization. Remarkably readable for a technically significant document, and historically important as an example of open research accelerating a competitive field. The decision to publish publicly rather than keep the approach proprietary changed the trajectory of the entire competition and, arguably, of recommendation research broadly.

3. Covington, P., Adams, J., & Sargin, E. (2016). "Deep Neural Networks for YouTube Recommendations." ACM Conference on Recommender Systems, 191-198. YouTube's engineering team describes the deep learning approach they use for their recommendation system. This paper reveals the two-stage architecture, the multi-objective training approach, and the feature engineering that characterizes modern large-scale recommendation systems. Essential reading for understanding how industrial systems depart from academic models.

4. Naumov, M., et al. (2019). "Deep Learning Recommendation Model for Personalization and Recommendation Systems." Facebook Research Technical Report. Facebook's description of DLRM (Deep Learning Recommendation Model), the architecture that powers Facebook and Instagram recommendations. Together with the YouTube paper, this gives a detailed view of how the two largest Western social media platforms approach recommendation at scale.

Training Objectives and Their Consequences

5. Chaslot, Guillaume (2019). "How Algorithms Can Learn to Discredit the Media." Medium / Partnership on AI. Guillaume Chaslot worked on YouTube's recommendation system until 2013 and became one of the most prominent public critics of engagement-based recommendation. This piece explains, from a practitioner perspective, how watch time optimization creates incentives that systematically push recommendations toward more extreme content. Essential for connecting technical architecture to real-world consequences.

6. Goodhart, C.A.E. (1975). "Problems of Monetary Management: The U.K. Experience." Papers in Monetary Economics, Reserve Bank of Australia. The original articulation of Goodhart's Law — "when a measure becomes a target, it ceases to be a good measure" — in the context of monetary policy. Reading the original source makes clear how general and how old this insight is, long predating its application to recommendation systems. Relevant for understanding why the proxy metric problem is not a recent discovery.

7. Haugen, Frances (2021). Senate Commerce Committee Testimony and Supporting Documents. Frances Haugen's testimony to the U.S. Senate and the accompanying documents (the "Facebook Papers") provide primary source evidence for the consequences of Facebook's "Meaningful Social Interactions" metric change. The documents show internal Facebook research demonstrating that the MSI update increased engagement with divisive content — evidence of Goodhart's Law operating at massive scale.

Feedback Loops and Algorithmic Effects

8. Pariser, Eli (2011). The Filter Bubble: What the Internet Is Hiding from You. Penguin Press. The book that introduced the "filter bubble" concept to mainstream discourse. While the technical understanding of recommendation systems has advanced considerably since 2011, Pariser's analysis of the political and epistemic consequences of personalization remains relevant and accessible. Best read alongside more recent empirical research that qualifies some of his stronger claims.

9. Sunstein, Cass (2017). #Republic: Divided Democracy in the Age of Social Media. Princeton University Press. Political scientist Sunstein examines the consequences of personalization for democratic deliberation and civic epistemics. Provides a political theory framework for understanding why filter bubbles matter beyond individual user experience. Engages with the empirical evidence on polarization and algorithmic media with more care than many popular treatments.

10. Jiang, R., et al. (2019). "Degenerate Feedback Loops in Recommender Systems." Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society. Academic analysis of how feedback loops in recommendation systems create self-reinforcing dynamics that harm users and distort content ecosystems. Includes formal modeling of preference amplification and popularity bias mechanisms. More technically demanding than other items on this list but provides the most rigorous treatment of feedback loop pathologies.

The Netflix Prize and Competition Research

11. Lohr, Steve (2009). "Netflix Awards $1 Million Prize and Starts a New Contest." New York Times, September 21, 2009. Contemporary reporting on the Netflix Prize conclusion. Includes details about why Netflix declined to implement the winning solution and quotes from Netflix executives about the gap between competition optimization and production reality. Essential for understanding the objective mismatch problem from primary sources.

12. Narayanan, A., & Shmatikoff, V. (2008). "Robust De-anonymization of Large Sparse Datasets." IEEE Symposium on Security and Privacy. The academic paper demonstrating that the Netflix Prize dataset could be de-anonymized by linking with public IMDb data. A foundational paper in data privacy research that showed behavioral data cannot be meaningfully anonymized through simple identifier removal. Essential reading for understanding the privacy implications of behavioral data collection for recommendation research.

Cold Start and Exploration-Exploitation

13. Leskovec, J., Backstrom, L., & Kleinberg, J. (2009). "Meme-tracking and the Dynamics of the News Cycle." ACM SIGKDD, 497-506. Analysis of how information spreads across online media. Relevant to the cold start problem in news and current events recommendation: how does a new item get discovered and recommended when it lacks engagement history? Provides empirical grounding for understanding how new content propagates (or fails to propagate) through recommendation systems.

14. Li, L., Chu, W., Langford, J., & Schapire, R.E. (2010). "A Contextual-Bandit Approach to Personalized News Article Recommendation." WWW 2010. Seminal paper on applying contextual bandits to recommendation. Provides formal treatment of the exploration-exploitation tradeoff in recommendation systems. Demonstrates that Thompson Sampling-style approaches outperform epsilon-greedy and other simpler exploration strategies. Technical but accessible with undergraduate statistics background.

Wellbeing, Proxy Metrics, and Alternatives

15. Matz, S.C., et al. (2023). "The Potential of Generative AI for Personalized Persuasion at Scale." Nature Human Behaviour. Examines the implications of highly personalized AI systems for individual autonomy and social outcomes. Relevant to the question of what alternatives to engagement-based recommendation might look like, particularly approaches that incorporate wellbeing signals rather than pure engagement proxies.

16. Bjornali, C., & Kaur, M. (2022). "Measuring User Wellbeing in Digital Platforms: A Framework for Operationalization." Proceedings of the ACM CHI Conference on Human Factors in Computing Systems. One of the few serious academic treatments of how wellbeing outcomes might be measured and incorporated into recommendation system design. Presents a framework for operationalizing wellbeing concepts in ways that could, in principle, be used as training objectives. Honest about the significant challenges involved.

17. Stray, J., et al. (2022). "Aligning AI With Shared Human Values." AAAI Workshop on AI Safety. Examines the general problem of specifying training objectives for AI systems that align with human values rather than with behavioral proxies. The analysis applies directly to recommendation systems and frames the proxy metric problem as an instance of the broader AI alignment challenge. Accessible and important for connecting the specific problems of recommendation systems to the broader field of AI ethics.

18. Roth, Yoel (2019). "The Authenticity of Non-Algorithmic Discovery." Internal Twitter presentation, later published. Yoel Roth, Twitter's former Head of Site Integrity, examines the philosophical and practical implications of distinguishing "organic" discovery (users finding content themselves) from algorithmic recommendation. Raises important questions about what we are actually trying to preserve when we worry about algorithmic influence on discovery, and what alternatives might look like.