Data Skeptic

Social Choice for Fair Recommendations

43 min

This episode of Data Skeptic explores the challenge of incorporating multiple fairness dimensions into recommender systems. Professor Robin Burke introduces 'Scruff D,' a novel social-choice-based framework that represents each fairness concern as a separate 'agent' which then competes or votes to influence the final recommendation list. The discussion contrasts this approach with traditional single-metric fairness methods and highlights the importance of dynamic, multi-objective fairness in real-world applications like news and micro-lending.

Summarized by Podsumo

🎧 Listen 🎙️ Ask about this episode

✨ Key Takeaways

1

Fairness is not a single metric; real-world systems must juggle multiple, often conflicting fairness objectives (e.g., gender representation, language diversity, genre balance) simultaneously.
2

The Scruff D framework decouples fairness from the core recommender by using a two-stage process: allocation (deciding which fairness agents are active) and preference aggregation (voting on the final ranked list).
3

A key limitation of re-ranking approaches is that if the original recommender is too biased (e.g., only recommends popular items), there are no diverse items to promote – ‘you can’t promote something if you don’t have it.’
4

The podcast contrasts Scruff D with regularization-based fairness methods, noting that Scruff D allows fairness objectives to be changed without retraining the entire model, offering greater flexibility for dynamic systems.
5

Future work focuses on moving beyond simple ranked lists to allocate items across complex, multi-element displays (e.g., carousels) and on treating content providers as first-class users of the recommender system.

💬 Notable Quotes

"“Fairness is not a computational concept. You can ask any five-year-old about whether they think what their brother got was fair or not. It’s part of human nature. But it’s also something that’s really hard to define. The best we can do is look at very specific cases.” — Robin Burke"
"“If your recommender has a lot of popularity bias, you have nothing to re-rank. You can’t promote something if you don’t have it.” — Robin Burke"
"“You kind of want to know, if I’m making this change, how is that going to have these ecosystem effects? If we had better tools to model recommendation ecosystems, we could simulate before we roll things out.” — Robin Burke"