Matching Markets: Managing Scale and Accuracy
summary
Competitive-equilibrium models are used to price and assign billions of items by compressing the value matrix via compact market via clustering or low-rank approximation, solving the small convex problem, then lifting prices and allocations back to the full market with provable bounds showing little welfare/regret loss in simulations. Edge weights and feed parameters are calibrated by combining biased simulators with a few online experiments and Bayesian optimization with Gaussian processes (expected improvement), using open-source tools Ax and BoTorch to converge to better feed and matching metrics.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:20Matching Markets
- 1:50Managing Scale
- 3:30Flow Problem
- 3:55In Practice
- 7:15Market Equilibria
- 9:22Tradeoffs
- 12:05Example
- 14:10Goals
- 14:35Blood Donations
- 15:20Fund Raisers
- 16:05Mentorship
- 17:09Facebook Dating
- 18:34Job Applications
- 21:11Learning Mode
- 39:40Optimization
- 43:21Conclusions
- 44:08Recap
related talks
26mA Game-Theoretic Perspective on the Recommender (Eco-)System
41mScaling Products to Millions of Customers
23mML-Powered Search & Recommendation 101
15mAI-Driven Application Optimization
48mRevisiting Performance Budgets
48mGrowing & Thriving in a Multi-model World
15mDevelop at Idea Velocity
01h 00m7 Years of DDD: Tackling Complexity in Large-Scale Marketing Systems