Research map / Learning theory and algorithms / Submodular and game algorithms
Optimal Auction Design in the Joint Advertising
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.
Derives an optimal Myerson-style joint auction for single-slot joint advertising and proposes BundleNet, a RegretNet-style neural mechanism with bundle-level IC constraints for multi-slot joint ads.
Problem
Revenue-optimal, approximately incentive-compatible auction design for joint advertising, where an ad slot is allocated to a bundle of a retailer and a supplier rather than a single bidder.
Approach
Characterizes the optimal single-slot joint auction via virtual values of bundles and critical values, and for multi-slot settings enforces IC on bundles (sum of bundle-level ex-post regrets) instead of per-bidder regrets, using a bipartite-graph feature fusion network with doubly stochastic allocation and sigmoid-normalized payments under an augmented Lagrangian.
Key results
- In single-slot setting U with 2 bundles, BundleNet achieves revenue 0.5286 versus 0.5247 for the optimal Myerson-like mechanism and 0.3811 for RVCG, with IC violation 0.0006.
- In multi-slot setting U with 5 slots and 5 bundles, BundleNet reaches revenue 1.4982 versus 1.4972 for JRegNet and 0.8420 for RVCG, with IC violation below 0.001.
- In multi-slot setting U with 12 bundles, BundleNet reaches 2.6495 versus 2.5868 for RVCG and 1.9973 for JRegNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BundleNet instead of the theoretical Optimal mechanism | Single-slot joint auction · U(0,1), U2–U5 | revenue: 0.5286 vs 0.5247; 0.6681 vs 0.6705; 0.7805 vs 0.7826; 0.8802 vs 0.8819 | Mixed |
| Use BundleNet instead of the theoretical Optimal mechanism | Single-slot joint auction · Truncated exponential, E2–E5 | revenue: 0.4248 vs 0.4249; 0.5460 vs 0.5479; 0.6354 vs 0.6470; 0.7215 vs 0.7376 | Hurt |
| Use BundleNet instead of RVCG | Multi-slot joint auction · Uniform U5x5–U12x5 | revenue: 1.4982 vs 0.8420; 1.7162 vs 1.2854; 1.9233 vs 1.6142; 2.0890 vs 1.8831; 2.2210 vs 2.0991; 2.4047 vs 2.2800; 2.5649 vs 2. | Helped |
| Use BundleNet instead of JRegNet | Multi-slot joint auction · Uniform U5x5–U12x5 | revenue: 1.4982 vs 1.4972; 1.7162 vs 1.6849; 1.9233 vs 1.8244; 2.0890 vs 1.9350; 2.2210 vs 1.9804; 2.4047 vs 1.9622; 2.5649 vs 1. | Mixed |
Surprising findings
- JRegNet attains higher single-slot revenue than the optimal mechanism in setting U with 2 bundles (0.5622 vs 0.5247), which the authors attribute to IC violation rather than superior design.
- JRegNet's multi-slot revenue plateaus and then declines as bundle count grows (e.g., setting U from 1.8244 at 7 bundles to 1.9973 at 12, and in setting LN from 3.4339 at 9 bundles down to 3.3848 at 12 and in setting N from 2.5185 at 9 down to 2.3272 at 12 bundles), while BundleNet keeps rising.
Limitations
- Multi-slot IC is only approximate (IC violation below 0.001 empirically) rather than guaranteed.
- Experiments use synthetic valuation distributions; real-world evaluation is limited to a cited dataset with unspecified details.
- Single-slot optimality theorem assumes regular bidders and a single slot with CTR 1 in experiments.
Experimental setup
Datasets: synthetic valuation samples from U(0,1), truncated exponential, truncated normal, truncated lognormal real-world dataset (cited, unspecified)
Benchmarks: synthetic single-slot settings U, E, N synthetic multi-slot settings U, LN, N (5 slots, 5-12 bundles)
Metrics: revenue empirical ex-post regret (IC violation)
Baselines: RVCG JRegNet Optimal joint auction (Myerson-like) VCG
Builds on: RegretNet JRegNet Myerson optimal auction RVCG JAMA
Closest papers by meaning
- Simultaneous Optimization of Bid Shading and Internal Auction for Demand-Side Platforms (AAAI 2024 · similarity 0.88)
- On Designing the Optimal Integrated Ad Auction in E-commerce Platforms (AAAI 2025 · similarity 0.87)
- AuctionNet: A Novel Benchmark for Decision-Making in Large-Scale Games (NeurIPS 2024 · similarity 0.86)
- A Combinatorial-Bandit Algorithm for the Online Joint Bid/Budget Optimization of Pay-per-Click Advertising Campaigns (AAAI 2018 · similarity 0.86)
- LLM-Auction: Generative Auction towards LLM-Native Advertising (NeurIPS 2026 · similarity 0.86)
- BundleFlow: Deep Menus for Combinatorial Auctions by Diffusion-Based Optimization (NeurIPS 2025 · similarity 0.85)
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.