Research map / Label-efficient and robust learning / Recognition and long-tail
Decongestion by Representation: Learning to Improve Economic Welfare in Marketplaces
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 learning item representations (feature masks) that decongest markets by optimizing a differentiable proxy of social welfare learned from revealed consumer choices, with theoretical conditions for when decongestion improves welfare.
Problem
Congestion in online marketplaces where users, making choices under partial information and with fixed decentralized prices, compete for the same items, reducing social welfare; the platform can only control item representations (which features are shown), and true preferences are unobserved.
Approach
Decompose welfare into a selection term (prices of chosen items) minus a decongestion penalty (excess demand per item), replace unobserved values and choices with a learned choice-prediction model, and relax the combinatorial top-k feature mask via a multinomial distribution with Gumbel top-k so the proxy is differentiable and trainable end-to-end on revealed choices.
Key results
- On the d=12 MovieLens setting, the learned representations approach the welfare oracle, with relative gains over random masks peaking around k=8.
- Prediction-based baselines (price-pred, choice-pred) improve only slowly with mask size k and fail at small k.
- Performance degrades as prices become higher or noisier, but the approach maintains performance relative to baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DbR to learn feature representations instead of prediction-based price-pred or choice-pred | MovieLens-100k, d=12 · MovieLens-100k | welfare | Helped |
| Use DbR to learn feature representations instead of price-pred | MovieLens-100k, d=100 · MovieLens-100k | welfare | Helped |
| Optimize the selection-minus-decongestion proxy instead of decongestion alone | Synthetic markets · Synthetic data | welfare | Helped |
| Use prediction or selection alone, without accounting for supply and congestion (ablation) | Synthetic markets · Synthetic data | welfare | Hurt |
| Choose a decongesting solution without the selection term (ablation) | Synthetic markets · Synthetic data | welfare | Mixed |
| Increase preference similarity (ablation) | Synthetic markets · Synthetic data | welfare | Hurt |
| Increase perceptive distortion in representations (ablation) | Synthetic markets · Synthetic data | welfare | Hurt |
| Contract the higher-value regime by using smaller valuation powers ρ (ablation) | Synthetic markets · Synthetic data | effectiveness of DbR | Hurt |
| Increase the mask size k (ablation) | DbR · MovieLens-100k, d=12 | relative welfare gain over random | Helped |
| Increase the mask size k (ablation) | DbR · MovieLens-100k, d=100 | welfare | Helped |
| Use choice-pred instead of DbR at k ≈ 50 (ablation) | MovieLens-100k, d=100 · MovieLens-100k | welfare | Mixed |
| Use higher or noisier prices instead of value-reflective CE prices (ablation) | DbR · MovieLens-100k, d=12, k=6 | welfare | Hurt |
Surprising findings
- Predictive accuracy and selection-only objectives do poorly: prediction-based masks can induce harmful congestion because they account only for demand, not supply.
Limitations
- Assumes fixed prices and unit supply, with prices not adapting to representations.
- Restricted to feature-subset representations and a linear non-negative user preference model; soft congestion in digital goods not addressed.
- Sufficient conditions for welfare improvement are simple and may be restrictive.
Experimental setup
Datasets: MovieLens-100k Yelp
Benchmarks: MovieLens-100k Yelp synthetic markets
Metrics: social welfare relative welfare gain over random masks number of allocated items
Baselines: price-pred choice-pred oracle welfare maximization random masks
Builds on: Gumbel top-k trick Adam
Closest papers by meaning
- How to Price Data: A Market Equilibrium Based Approach (ICML 2026 · similarity 0.86)
- Learning Classifiers That Induce Markets (ICML 2025 · similarity 0.85)
- Equilibrium of Data Markets with Externality (ICML 2024 · similarity 0.85)
- Multidimensional Bayesian Utility Maximization: Tight Approximations to Welfare (NeurIPS 2025 · similarity 0.84)
- A Market for Accuracy: Classification Under Competition (ICML 2025 · similarity 0.84)
- A Parametric Contextual Online Learning Theory of Brokerage (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.