Research map / Learning theory and algorithms / Submodular and game algorithms
Beyond Self-Interest: How Group Strategies Reshape Content Creation in Recommendation Platforms?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies how creator group strategies in a bandit-style content creation game with top-K exposure rewards change equilibria and user welfare, and shows engagement rewards bound the price of anarchy.
Problem
Creators in recommendation platforms may form groups (e.g., under an MCN) that coordinate to maximize collective utility, which can shift content distribution and reduce user welfare relative to self-interested creators.
Approach
Extends the C3 content creation game to group behavior: creators in a group jointly maximize group utility; shows small groups have limited effect, large groups can cause welfare loss under exposure rewards, and engagement rewards bound the PoA (≤K+1 general, ≤2 binary case).
Key results
- Under exposure rewards the PoA can be arbitrarily large (Theorem 5.3), and a single group of all creators can cause user welfare loss as large as O(n) in the TvN game.
- With engagement rewards, PoA ≤ K+1 in general (β≥1, constant attention) with a matching-order lower bound of 1+(K-1)e^{-1/β}; PoA ≤2 for binary relevance and attention with β→0.
- In simulations (n=30), user welfare under exposure rewards drops sharply when group size reaches 25 or 30, while engagement rewards keep welfare stably high; engagement-reward PoA stays close to 1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace exposure-based rewards with user engagement rewards | Creator groups in C3 recommendation games · General C3 games | PoA: PoA →∞; PoA ≤ 1 + K | Helped |
| Use engagement rewards rather than exposure rewards | Creators in synthetic simulations · Synthetic simulation, group size 30, α = 1.1 | user welfare: 3.26±0.02 -> 4.77±0.01 | Helped |
| Increase group size from 10 to 30 under exposure rewards | Creator group in synthetic simulations · Synthetic simulation, α = 1.1 | user welfare: 4.72±0.04 -> 3.26±0.02 | Hurt |
| Increase K beyond the attention cutoff τ | Top-K recommendation with group behavior · TvN game with diminishing attention | user welfare | Hurt |
| Increase β from 0.1 to 10 | Creators under constant attention scores · C3 simulation with K = n | user welfare: β = 0.1; β = 10 | Helped |
| Replace exposure rewards with engagement rewards | Creator groups in binary-relevance C3 games · Binary relevance and attention scores | PoA: PoA ≤ 2 | Helped |
Surprising findings
- Under diminishing attention, user welfare does not necessarily increase with larger K or β when creators group together.
- Group behavior has limited effect on equilibrium when group size is small (Theorem 4.6).
Limitations
- Theoretical analysis restricted to bandit C3 games with orthogonal user/creator embeddings in the main results.
- Simulations use a synthetic environment with a small number of users and creators (n=30, m=10).
- Worst-case PoA under engagement rewards with constant attention still grows with K; mitigation relies on intervention methods not fully analyzed here.
Experimental setup
Metrics: user welfare price of anarchy
Baselines: individual (vanilla) creator equilibrium exposure reward mechanism
Builds on: Content Creation Competition (C3) game Price of Anarchy Local Better Response (LBR) Yao et al. 2024 user importance reweighting
Closest papers by meaning
- Policy Design for Two-sided Platforms with Participation Dynamics (ICML 2025 · similarity 0.88)
- Recommender Ecosystems: A Mechanism Design Perspective on Holistic Modeling and Optimization (AAAI 2024 · similarity 0.87)
- Unveiling User Satisfaction and Creator Productivity Trade-Offs in Recommendation Platforms (NeurIPS 2024 · similarity 0.86)
- The Burden of Interactive Alignment with Inconsistent Preferences (NeurIPS 2025 · similarity 0.85)
- Private Blotto: Viewpoint Competition with Polarized Agents (AAAI 2025 · similarity 0.85)
- User-Creator Feature Polarization in Recommender Systems with Dual Influence (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.