Data Skeptic

Kyle Polich sits down with Yashar Deldjoo, research scientist and Associate Professor at the Polytechnic University of Bari, to explore how recommender systems have evolved and why trustworthiness matters. They unpack key dimensions of responsible AI, including robustness to adversarial attacks, privacy, explainability, and fairness, and discuss how LLMs introduce new risks like hallucinations.
The episode closes with a look at "agentic" recommender systems, where tools and memory shift recommendations from ranked lists to end-to-end task completion.

Goodreads star ratings can be misleading as measures of "book quality," and research from Hannes Rosenbusch suggests that for many professionally published books, differences between readers often matter more than differences between books. The episode also explores how to model reader preferences, why reviews often reveal more about the reviewer than the text, and how LLMs can aid computational literary research while still falling short of human editors in creative writing.

Ervin Dervishaj, a PhD student at the University of Copenhagen, discusses his research on disentangled representation learning in recommender systems, finding that while disentanglement strongly correlates with interpretability, it doesn't consistently improve recommendation performance. The conversation explores how disentanglement acts as a regularizer that can enhance user trust and interpretability at the potential cost of some accuracy, and touches on the future of large language models in denoising user interaction data.

Ekaterina (Kat) Fedorova from MIT EECS joins us to discuss strategic learning in recommender systems—what happens when users collectively coordinate to game recommendation algorithms. Kat's research reveals surprising findings: algorithmic "protest movements" can paradoxically help platforms by providing clearer preference signals, and the challenge of distinguishing coordinated behavior from bot activity is more complex than it appears. This episode explores the intersection of machine learning and game theory, examining what happens when your training data actively responds to your algorithm.

Anas Buhayh discusses multi-stakeholder fairness in recommender systems and the S'mores framework—a simulation allowing users to choose between mainstream and niche algorithms. His research shows specialized recommenders improve utility for niche users while raising questions about filter bubbles and data privacy.