Research map / Label-efficient and robust learning / Recognition and long-tail
User-Creator Feature Polarization in Recommender Systems with Dual Influence
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.
Defines user-creator feature dynamics for recommender systems with dual influence on users and creators, proves that under non-zero recommendation probabilities the system polarizes, and finds top-k truncation can mitigate polarization while diversity-boosting does not.
Problem
Recommender systems influence both users' preferences and creators' content, so diversity may erode over time; existing diversity-boosting methods assume static users and content, and it is unclear whether they still work under this dual influence.
Approach
A dynamical model in which users and creators are unit-norm embeddings updated toward the items/users they interact with (softmax recommendation, biased-assimilation user update, creator update toward recommended users' preferences); proves almost-sure convergence to consensus or bi-polarization when all creators have nonzero recommendation probability, and shows top-k truncation can instead prese
Key results
- Top-k truncation with k=1 on synthetic data reduces tendency to polarization to 0.27 (from 1.00 without truncation at k=50, beta=1) while raising creator diversity to 1.40.
- On MovieLens 20M with beta=1, top-k truncation at k=1 lowers tendency to polarization from 1.00 (k=2000) to 0.20 and raises creator diversity to 1.40, but drives recommendation diversity to 0.
- Under softmax and uniform-traffic recommendation with non-zero probabilities, the theory proves the dynamics reach consensus or bi-polarization almost surely; larger beta reduces tendency to polarization in simulations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase recommendation sensitivity β (ablation) | User-creator feature dynamics · Synthetic data and MovieLens 20M | Tendency to Polarization | Helped |
| increase creator update rate η_c (ablation) | User-creator feature dynamics · Synthetic data | Creator Diversity | Hurt |
| increase user update rate η_u (ablation) | User-creator feature dynamics · Synthetic data | Tendency to Polarization | Hurt |
| increase the number of fixed user-feature dimensions (ablation) | User-creator feature dynamics · Synthetic data | Tendency to Polarization | Helped |
| add diversity-aware recommendation boosting (increase ρ) (ablation) | MovieLens 20M dynamics · MovieLens 20M | Tendency to Polarization | Hurt |
| apply top-k truncation, reducing k (ablation) | Softmax recommender · Synthetic data and MovieLens 20M | Creator Diversity | Helped |
| apply top-k truncation, reducing k (ablation) | Softmax recommender · Synthetic data and MovieLens 20M | Recommendation Diversity | Hurt |
| set threshold truncation to τ=0 (ablation) | Softmax recommender · Synthetic data and MovieLens 20M | Tendency to Polarization | Hurt |
| raise threshold truncation to τ=0.707 (ablation) | Softmax recommender · Synthetic data and MovieLens 20M | Tendency to Polarization | Helped |
Surprising findings
- Diversity-boosting objectives and uniform random traffic, which improve diversity in static settings, make polarization more likely under dual influence.
- Efficiency-oriented top-k truncation can alleviate polarization and improve creator diversity, while reducing recommendation diversity.
Limitations
- Theorem requires update rates eta_c <= eta_u L_f / 2 and eta_u < 1/2, and the proof assumes the sign function for creator impact g; the inner-product user impact f violates the |f|>=L_f assumption in experiments.
- Convergence rate to polarization is not characterized and is left as an open question.
- Experiments use simulated dynamics on synthetic data and MovieLens 20M, not live platform behavior.
Experimental setup
Datasets: MovieLens 20M synthetic unit-sphere features
Benchmarks: synthetic simulation MovieLens 20M
Metrics: creator diversity recommendation diversity recommendation relevance tendency to polarization
Baselines: diversity-aware objective (relevance plus diversity reward) uniform traffic threshold truncation
Builds on: opinion dynamics (bounded confidence / consensus models) performative prediction
Closest papers by meaning
- Unveiling User Satisfaction and Creator Productivity Trade-Offs in Recommendation Platforms (NeurIPS 2024 · similarity 0.86)
- Optimizing Social Network Interventions via Hypergradient-Based Recommender System Design (ICML 2025 · similarity 0.85)
- Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · similarity 0.85)
- Recommender Ecosystems: A Mechanism Design Perspective on Holistic Modeling and Optimization (AAAI 2024 · similarity 0.85)
- Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommender Systems (NeurIPS 2026 · similarity 0.84)
- Beyond Self-Interest: How Group Strategies Reshape Content Creation in Recommendation Platforms? (ICML 2025 · 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.