Research map / Reinforcement learning / Visual and representation RL
Policy Design for Two-sided Platforms with Participation Dynamics
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 viewer/provider participation dynamics on two-sided platforms, shows myopic-greedy recommendation can be suboptimal under heterogeneous population effects, and proposes a look-ahead policy (interpolated with myopic-greedy) that optimizes welfare at projected long-term populations.
Problem
Recommendation policies on two-sided platforms ignore how viewer and provider population growth feeds back into long-term welfare, so myopic-greedy policies can cause provider loss and lower long-run viewer welfare.
Approach
Models viewer and provider population dynamics with population-dependent quality (population effects), analyzes stability as a game's Nash equilibria, decomposes regret into population and policy regret, and proposes a look-ahead policy that optimizes welfare at the projected reference population, interpolated with the myopic-greedy policy.
Key results
- In the synthetic experiment, the look-ahead policy (beta=1.0) increases provider populations while myopic-greedy decreases them, yielding higher total welfare than myopic-greedy and uniform.
- On KuaiRec, an interpolated policy with beta=0.6 performed best among compared methods, and beta=1.0 maintained high welfare with provider population close to uniform.
- Theory shows myopic-greedy is optimal under linear, homogeneous population effects, but can be suboptimal under heterogeneous effects (counterexample proposition).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the look-ahead policy instead of myopic-greedy | Two-sided platform with heterogeneous concave population effects · Synthetic experiment | total welfare | Helped |
| Use the look-ahead policy instead of myopic-greedy | Two-sided platform with heterogeneous concave population effects · Synthetic experiment | provider population | Helped |
| Use the look-ahead policy instead of uniform random | Two-sided platform with heterogeneous concave population effects · Synthetic experiment | total welfare | Helped |
| Use uniform random instead of myopic-greedy | Two-sided platform with heterogeneous concave population effects · Synthetic experiment | provider population | Helped |
| Set interpolation weight β to 1.0 (pure look-ahead) (ablation) | Look-ahead policy · Synthetic experiment | total welfare | Helped |
| Use look-ahead instead of myopic-greedy | Two-sided platform simulated from KuaiRec · KuaiRec simulation | total welfare | No clear effect |
| Use myopic-greedy instead of uniform random | Two-sided platform simulated from KuaiRec · KuaiRec simulation | total welfare | Helped |
| Set interpolation weight β to 0.6 (ablation) | Look-ahead and myopic-greedy interpolation · KuaiRec simulation | total welfare | Helped |
| Set interpolation weight β to 1.0 (pure look-ahead) (ablation) | Look-ahead and myopic-greedy interpolation · KuaiRec simulation | population regret | Helped |
| Set interpolation weight β to 1.0 (pure look-ahead) (ablation) | Look-ahead and myopic-greedy interpolation · KuaiRec simulation | policy regret | Hurt |
Surprising findings
- Uniform random policy can outperform myopic-greedy in the long term when population effects are large.
- On KuaiRec, myopic-greedy outperformed uniform random, unlike the synthetic setting.
Limitations
- Analysis of optimality is restricted to linear, homogeneous population effects; general case only has counterexamples.
- Real-data experiment uses simulated subgroups and population effects derived from permutations, not real population dynamics.
- Requires estimating dynamics functions, which adds approximation error.
Experimental setup
Datasets: KuaiRec
Benchmarks: KuaiRec (simulated subgroup environment) synthetic two-sided simulation
Metrics: total viewer welfare provider population cumulative regret population regret policy regret
Baselines: myopic-greedy policy uniform random policy
Builds on: performative prediction Cournot duopoly game neural collaborative filtering
Closest papers by meaning
- Beyond Self-Interest: How Group Strategies Reshape Content Creation in Recommendation Platforms? (ICML 2025 · similarity 0.88)
- Recommender Ecosystems: A Mechanism Design Perspective on Holistic Modeling and Optimization (AAAI 2024 · similarity 0.86)
- Private Blotto: Viewpoint Competition with Polarized Agents (AAAI 2025 · similarity 0.84)
- Reaching a Consensus in Predictive Loops (NeurIPS 2026 · similarity 0.83)
- Collective Counterfactual Explanations: Balancing Individual Goals and Collective Dynamics (NeurIPS 2025 · similarity 0.83)
- Policy Design in Long-run Welfare Dynamics (ICLR 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.