Research map / Learning theory and algorithms / Prediction and decision losses
Learning Classifiers That Induce Markets
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 market-aware strategic classification in which a linear classifier induces a competitive market for features, with exact and differentiable price computation and a market hinge loss for end-to-end learning.
Problem
Strategic classification assumes feature-manipulation costs are exogenous and fixed, but in practice deploying a classifier creates demand for important features, so costs emerge from a market that the classifier itself induces.
Approach
Model feature sellers as competing firms with linear costs; the classifier's weight vector determines demand, and equilibrium prices are proportional to w. Users with budgets buy features to cross the boundary. Provide an exact O(m log m) price algorithm, a differentiable smooth-price relaxation, and a market hinge loss so learning can anticipate the induced market.
Key results
- On Adult with budget inequity matching the original data, MASC outperforms the naive classifier by a large margin and outperforms strat in the long term for scales above 2^4.
- Synthetic example: a market-aware classifier using only an uninformative feature reaches 0.96 accuracy, while the naive classifier reaches 0.29.
- With uniform budgets no threshold exceeds 55% accuracy on separable data, but negative correlation between budget and distance enables accuracy near 0.95.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the market mechanism with x2 instead of class-informative x1 | Market-aware classifier · Synthetic 2-feature market | accuracy: 0.29 to 0.96 | Helped |
| Use uniform budgets while varying the decision threshold | Market-aware threshold classifier · Synthetic separable data | accuracy: no threshold obtains accuracy above 55% | Hurt |
| Make budgets increase linearly with z, from bmin=1 to bmax=5 | Market-aware threshold classifier · Synthetic separable data | accuracy: ∼0.5 to ∼0.95 | Helped |
| Increase budget-inequality scale | MASC · Adult | test accuracy | Helped |
| Account for market price adaptation during classifier learning | MASC vs. naive classifier · Adult | test accuracy | Helped |
| Account for market price adaptation during classifier learning | MASC vs. strategic classifier · Adult, long-term performance | test accuracy | Helped |
| Evaluate the strategic classifier after prices re-equilibrate rather than at current prices | Strategic classifier · Adult | test accuracy | Hurt |
| Use smoothed market prices instead of exact empirical prices | Market hinge loss · Synthetic 1D loss landscape | loss-curve smoothness | Helped |
Surprising findings
- Raising the classification threshold can cause more points to cross, not fewer, because prices adapt.
- Classifiers can exploit features uninformative of the label because the induced market makes them separable.
Limitations
- Stylized market model with perfect information, unlimited supply, fixed exclusive sellers and no externalities.
- Real-data experiments use simulated market behavior and synthetic budgets derived from capital_gain.
- Learning can drive accuracy toward ~50% after market formation for high-accuracy pre-market classifiers when budgets are uncorrelated with distance.
Experimental setup
Datasets: Adult Folktables
Benchmarks: Adult Folktables
Metrics: accuracy welfare social burden fraction of positives moving fraction of negatives crossing
Baselines: naive classifier strat (fixed-price strategic classifier)
Builds on: strategic classification strategic hinge loss linear costs in strategic classification
Closest papers by meaning
- A Market for Accuracy: Classification Under Competition (ICML 2025 · similarity 0.88)
- Classification Under Strategic Self-Selection (ICML 2024 · similarity 0.86)
- Decongestion by Representation: Learning to Improve Economic Welfare in Marketplaces (ICLR 2024 · similarity 0.85)
- Incentivizing Desirable Effort Profiles in Strategic Classification: The Role of Causality and Uncertainty (NeurIPS 2025 · similarity 0.85)
- Pricing and Competition for Generative AI (NeurIPS 2024 · similarity 0.84)
- Strategic Classification with Non-Linear Classifiers (NeurIPS 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.