Creator Incentives in Recommender Systems: A Cooperative Game-Theoretic Approach for Stable and Fair Collaboration in Multi-Agent Bandits

A team of researchers from NYU and Meta AI has published a groundbreaking paper titled 'Creator Incentives in Recommender Systems: A Cooperative Game-Theoretic Approach for Stable and Fair Collaboration in Multi-Agent Bandits,' accepted as an Oral Presentation at AISTATS 2026. The work tackles a fundamental problem in modern platforms: how to fairly allocate credit and rewards among content creators when an AI recommender system's learning is interdependent. User feedback on one creator's content influences the system's model, which in turn affects the exposure and success of other creators. The authors formalize this as a multi-agent stochastic linear bandit problem, where a coalition of creators' collective value is defined as the negative sum of their cumulative regrets—essentially, how much better they could have done.

They apply cooperative game theory with transferable utility (TU) to this setup. A key finding is that for identical creators, the induced TU game is convex under mild conditions, guaranteeing a non-empty 'core'—a set of stable payoff allocations where no subgroup has incentive to break away. This core contains the Shapley value, a classic fairness concept. For the more realistic case of heterogeneous creators, convexity isn't guaranteed, so the team proposes a practical, regret-based payout rule. This rule satisfies three of the four Shapley axioms (efficiency, symmetry, and dummy player) and crucially, still lies within the core, ensuring coalitional stability.

The framework was tested on the MovieLens-100k dataset, illustrating scenarios where empirical payouts to creators align with—and diverge from—the theoretically fair Shapley value across different algorithmic settings. This provides a rigorous, mathematical foundation for platform economics, moving beyond simple metrics like view counts to account for the network effects and collaborative learning inherent in AI-driven systems. It offers a tool for platforms to design incentive structures that are both stable (preventing creator exodus) and perceived as fair, which is critical for long-term ecosystem health.