Research map / Reinforcement learning / Visual and representation RL
Unveiling User Satisfaction and Creator Productivity Trade-Offs 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.
Proposes the Cournot Content Creation Competition (C^4) game to model how a relevance/exploration-controlled recommender affects creators' production volume, proves a unique Nash equilibrium and a user-satisfaction vs. content-volume trade-off, and optimizes per-user exploration strength via approximated gradient descent.
Problem
Recommendation algorithms on UGC platforms shape creators' production frequency, but existing creator-competition models ignore content creation volume; the trade-off between short-term user satisfaction and long-term content supply is unclear.
Approach
Model creators as choosing production quantity in a Cournot-style game where traffic is allocated by a personalized probabilistic matching (softmax over noisy relevance scores with per-user exploration strength beta); prove a unique PNE, show user satisfaction increases and content volume decreases in beta, and optimize beta via implicit-differentiation gradients with Sherman-Morrison-Woodbury ske
Key results
- Optimized per-user beta improves welfare W_lambda by over 20% relative to exact matching in both synthetic and MovieLens environments.
- Theory: U is strictly increasing and V is strictly decreasing in beta (for large beta, m=1), showing a trade-off between user satisfaction and content volume.
- Personalized beta yields slightly better welfare than homogeneous beta in both environments.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase homogeneous matching precision by raising β | C⁴ creator competition · Single-user setting (m=1) | user satisfaction U | Helped |
| increase homogeneous matching precision by raising β | C⁴ creator competition · Single-user setting (m=1), β ≥ β₀ | content creation volume V | Hurt |
| increase homogeneous β | C⁴ creator competition · Synthetic and MovieLens simulations | user satisfaction U | Helped |
| increase homogeneous β | C⁴ creator competition · Synthetic and MovieLens simulations | content creation volume V | Hurt |
| optimize personalized β with approximated gradient descent instead of exact matching for all users | PPM recommendation policy, λ=0.5 · Synthetic and MovieLens environments | social welfare Wλ: over 20% | Helped |
| use personalized β rather than homogeneous β | PPM recommendation policy, λ=0.5 · Synthetic and MovieLens environments | social welfare Wλ | Helped |
| assign smaller β (more exploration) to less selective users | Optimized personalized PPM policy · Synthetic and MovieLens environments | social welfare Wλ | Helped |
Surprising findings
- Increasing matching accuracy (beta) raises user satisfaction but reduces total content creation volume, because creators with deterministic traffic allocation have little incentive to produce more.
Limitations
- Assumes creators keep a fixed niche and only vary production quantity
- Evaluation is purely simulation-based with synthetic and MovieLens-derived creator/user populations
- Monotonicity proof for V is for m=1 with homogeneous population; broader settings are supported only empirically
Experimental setup
Datasets: MovieLens-1m
Benchmarks: synthetic clustered environment (m=1000 users, n=200 creators) MovieLens-1m simulated environment
Metrics: total user satisfaction U total content creation volume V social welfare W_lambda
Baselines: exact matching (beta=100 deterministic) homogeneous beta
Builds on: Content Creator Competition (C^3) Tullock contest Cournot competition multi-agent mirror descent
Closest papers by meaning
- Beyond Self-Interest: How Group Strategies Reshape Content Creation in Recommendation Platforms? (ICML 2025 · similarity 0.86)
- User-Creator Feature Polarization in Recommender Systems with Dual Influence (NeurIPS 2024 · similarity 0.86)
- Recommender Ecosystems: A Mechanism Design Perspective on Holistic Modeling and Optimization (AAAI 2024 · similarity 0.84)
- On the Power of Strategic Corpus Enrichment in Content Creation Games (AAAI 2025 · similarity 0.82)
- Ready for You When You Are Back: Content-Driven Session-Based Recommendation for Continuity of Experience (AAAI 2025 · similarity 0.82)
- The Burden of Interactive Alignment with Inconsistent Preferences (NeurIPS 2025 · similarity 0.82)
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.