Overview of Social Choice for Fair Recommendations
This episode of Data Skeptic features Robin Burke, a professor at the University of Colorado Boulder and a longtime recommender systems researcher, discussing how recommender systems have evolved from early content, collaborative, and knowledge-based approaches into modern systems that must balance accuracy, fairness, and ecosystem effects. The centerpiece of the conversation is SCRUFF-D — Social Choice for Recommendation Under Fairness, Dynamic — a multi-agent framework that treats different fairness concerns as separate agents competing and cooperating to shape recommendation outputs over time.
Key Takeaways
- Recommender systems are no longer just about accuracy.
- Evaluation has expanded to include usability, actionability, journalistic value, provider fairness, and long-term ecosystem health.
- Fairness is multi-dimensional and dynamic.
- Burke argues that fairness cannot usually be reduced to one metric or one protected group.
- Real systems often need to account for multiple fairness goals simultaneously.
- SCRUFF-D reframes fairness as a social choice problem.
- Each fairness concern becomes an agent.
- The system first decides which agents are relevant in a given moment, then aggregates their preferences with the user’s recommendations.
- Recommendation systems affect entire ecosystems.
- Users, providers, artists, creators, lenders, and platforms all react to recommendations.
- Feedback loops can amplify bias over time if not actively managed.
How Recommender Systems Have Changed
Early recommender systems
Burke traces his work back to the early 1990s, when the term “recommender systems” did not yet exist. His background was in:
- Intelligent tutoring systems
- Case-based reasoning
- Knowledge-based recommendation
He explains that recommender systems originally emerged from several separate traditions:
- Content-based recommendation: uses item features
- Collaborative recommendation: uses patterns across users
- Knowledge-based recommendation: uses domain knowledge and constraints
His influential survey on hybrid recommender systems helped frame the idea that these methods should not be seen as mutually exclusive.
Modern recommender systems
Today, what used to be called “hybrid” is often just considered standard practice:
- Add as many useful features as possible
- Combine user, item, and context signals
- Model multiple sources of information in one system
Burke notes that the field has shifted from debating whether to combine methods to assuming that richer feature sets are the norm.
What Fairness Means in Recommender Systems
Fairness is hard to define
Burke emphasizes that fairness is not a purely computational concept. It is:
- Philosophical
- Context-dependent
- Stakeholder-dependent
- Often impossible to capture with a single formula
Examples of fairness concerns
He gives several examples of how fairness can vary by domain:
- Music recommendation
- Ensure women artists are not under-recommended
- Support non-English language artists
- Avoid over-promoting only the most popular performers
- News recommendation
- Balance what users want to click on with what they should know
- Support journalism’s role in keeping citizens informed
- Libraries and e-commerce
- Recommend items that are actually available
- Avoid recommending backordered or inaccessible items
- Micro-lending
- In Kiva’s case, fairness must support global financial inclusion across changing provider and borrower needs
Fairness must be measured carefully
Burke warns that observed user behavior may already be shaped by the recommender system itself, creating feedback loops that complicate fairness measurement.
SCRUFF-D: Social Choice for Recommendation Under Fairness, Dynamic
Core idea
SCRUFF-D treats each fairness concern as an agent with its own objective. Instead of trying to optimize one fairness metric at training time, the system:
- Allocates fairness agents to the current recommendation opportunity
- Aggregates the preferences of the user and the active agents
- Produces a final re-ranked recommendation list
Why this approach
The framework was designed to address limitations in prior fairness work:
- Most studies focus on one fairness dimension at a time
- Many assume the same fairness definition applies everywhere
- Static approaches do not handle changing fairness needs over time
Dynamic fairness
The “D” in SCRUFF-D reflects the fact that fairness needs shift over time. For example:
- A provider group may have been underexposed recently
- A particular category may need more attention now
- The system may need to respond to historical imbalance, not just one-shot fairness
How SCRUFF-D Works
1. Allocation
This step decides which fairness agents should be active for a given recommendation opportunity.
Possible strategies include:
- Lottery-style selection
- Randomly choose one relevant agent
- Weighted allocation
- Give different fairness concerns different levels of influence
- Compatibility-based selection
- Only activate agents relevant to the current user or context
2. Preference aggregation
Once the active agents are chosen, SCRUFF-D combines:
- The recommender’s ranking, representing user preferences
- The agents’ preferences, representing fairness goals
This is treated as a voting problem.
Possible aggregation methods discussed include:
- Score-based voting rules
- Simple score boosts
- Greedy sublist re-ranking
- Other social choice mechanisms
What the experiments showed
Burke notes two important findings:
- Sometimes allocation matters more than the voting rule
- In other cases, the aggregation method does affect fairness/accuracy trade-offs
The best choice depends on the dataset, recommender, and fairness objective.
Why Not Just Use Regularization?
The interviewer asks about an alternative approach: adding fairness directly into the training objective through regularization.
Burke’s response:
- Regularization can work, but it becomes difficult when:
- There are multiple fairness dimensions
- The fairness objectives are heterogeneous
- The system needs to adapt dynamically
- SCRUFF-D has an advantage because:
- It can be adjusted without retraining the recommender model
- Fairness definitions can be changed more flexibly
- It is easier to respond to changing system needs
A limitation of re-ranking
A key caveat is that a re-ranker can only improve fairness if the base recommender produces candidate items that can be promoted. If the initial model is too biased or too narrow, the re-ranker may have little to work with.
Future Directions
Burke identifies several promising directions for the field:
More realistic recommendation displays
Most recommender research still assumes a simple ranked list, but real interfaces are more complex:
- Multiple carousels
- Different screen regions with different purposes
- Nonlinear layouts
- Mixed content surfaces
He suggests that social-choice thinking could help generalize recommendation beyond top-10 lists to allocation across display real estate.
Better modeling of recommendation ecosystems
He argues for more research on how recommendation systems interact with:
- Users
- Providers
- Market dynamics
- Long-term fairness drift
Simulation tools could help predict how policy or algorithm changes affect the ecosystem before deployment.
Treat providers as first-class users
Burke also stresses that providers — such as artists, creators, and lenders — should be understood as users of the system too. They have:
- Visibility needs
- Explanation needs
- Fair access needs
This broader view moves recommender systems research beyond user clicks and toward full ecosystem design.
Notable Insights
- “Fairness is not a computational concept” — Burke emphasizes that fairness cannot be reduced to a single metric in the general case.
- Recommendation systems should be evaluated not just by accuracy, but by whether they create a healthy and sustainable ecosystem.
- A good fairness system must be dynamic, because past recommendations change the future state of the system.
- Multi-objective fairness is better handled by a framework that can represent multiple stakeholders simultaneously, rather than forcing everything into one loss function.
Links and Follow-Up
Burke says the best way to follow his work is through:
- Google Scholar
- Mastodon
- Bluesky
He notes that he posts infrequently, but publications are the most reliable way to track his research.