Research map / Learning theory and algorithms / Submodular and game algorithms
Simultaneous Optimization of Bid Shading and Internal Auction for Demand-Side Platforms
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.
Jointly optimizes bid shading and an ex-post incentive-compatible internal auction for a DSP in open first-price auctions by reducing the problem to learning monotone bid functions, implemented as a partially monotone min-max network trained with a sigmoid surrogate loss.
Problem
A demand-side platform (DSP) bidding in open first-price ad exchange auctions must shade bids without accurate winning-probability estimates, while allocating won impressions among advertiser clients via an internal auction that is ex-post incentive compatible and individually rational.
Approach
Proves that joint optimization of bid shading and ex-post IC/IR internal auction reduces, without loss of optimality, to choosing one monotone bid function per advertiser; implements these with a partially monotone MIN-MAX neural network and optimizes a sigmoid-smoothed surrogate loss end-to-end, without assuming a parametric winning-price distribution.
Key results
- On industrial iPinYou data, MINMAX achieves APO 0.5542 for surplus vs 0.5159 for Dist-L, and 0.3491 vs 0.2844 for profit (7.42% surplus lift and 22.75% profit lift over Dist-L).
- In online A/B test on a leading advertising platform, MINMAX improves surplus by 6.41% over Dist-L with only 1.80% impression increase.
- On synthetic uniform and exponential distributions, MINMAX's achieved objective is close to the optimum (e.g., 0.9997 and 1.0019 for surplus with uniform and exponential).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the Dist-L bid-shading approach with MINMAX | DSP bidding and internal auction · Industrial data, surplus maximization | APO: 0.5159(0.0010) -> 0.5542(0.0014); 7.42% surplus lift | Helped |
| Replace the Dist-L bid-shading approach with MINMAX | DSP bidding and internal auction · Industrial data, profit maximization | APO: 0.2844(0.0038) -> 0.3491(0.0020); 22.75% profit lift | Helped |
| Replace Dist-L with MINMAX for online bid shading | Real-time bid shading serving module · Online A/B test, surplus maximization | surplus: 6.41% | Helped |
| Replace Dist-L with MINMAX for online bid shading | Real-time bid shading serving module · Online A/B test | impressions: 1.80% | Mixed |
| Remove the feature-dependent slope term from the monotone bid function (ablation) | MINMAX · Log-normal, profit maximization | AO: 0.9441 -> 0.9242 | Hurt |
| Use a separate feature-dependent slope weight for each group (ablation) | MINMAX · Log-normal, profit maximization | AO: 0.9441 -> 0.9412 | Hurt |
| Set Q = Z = 1 (ablation) | MINMAX · Log-normal, profit maximization | AO: 0.9441 -> 0.9338 | Hurt |
| Set Q = Z = 1 (ablation) | MINMAX · Log-normal, profit maximization | AO: 0.9242 -> 0.9338 | Helped |
| Remove the feature-dependent slope term from the monotone bid function (ablation) | MINMAX · Gamma, profit maximization | AO: 0.9339 -> 0.9274 | Hurt |
| Use a separate feature-dependent slope weight for each group (ablation) | MINMAX · Gamma, profit maximization | AO: 0.9339 -> 0.9324 | Hurt |
| Set Q = Z = 1 (ablation) | MINMAX · Gamma, profit maximization | AO: 0.9339 -> 0.9297 | Hurt |
| Use Dist-L instead of MINMAX | Bid shading · Log-normal, surplus maximization | AO: 0.9928 -> 0.9954 | Helped |
Surprising findings
- Dist-L beats Dist-G on log-normal but loses on gamma data, illustrating that distribution-estimation methods are sensitive to distribution assumptions.
Limitations
- Relies on assumption of independent advertiser values given context and known feedback distribution structure for the theory
- Offline evaluation on synthetic data relies on known optimal solutions; real data uses an upper bound proxy
- Online experiments only for surplus maximization against a single baseline (Dist-L)
Experimental setup
Datasets: iPinYou RTB dataset synthetic data (log-normal, gamma, uniform, exponential distributions)
Benchmarks: iPinYou RTB dataset synthetic distributions online A/B test
Metrics: achieved objective (AO) bid deviation (BD) achieved post-objective (APO) surplus lift profit lift impression increase
Baselines: BSR Dist-S Dist-G Dist-L Mod1 Mod2
Builds on: MyersonNet Myerson optimal auction Neural Auction (Liu et al. 2021) DeepFM
Closest papers by meaning
- Optimal Auction Design in the Joint Advertising (ICML 2025 · similarity 0.88)
- Leveraging the Hints: Adaptive Bidding in Repeated First-Price Auctions (NeurIPS 2022 · similarity 0.88)
- AuctionNet: A Novel Benchmark for Decision-Making in Large-Scale Games (NeurIPS 2024 · similarity 0.88)
- Optimal Auction Design for Mixed Bidders (AAAI 2025 · similarity 0.87)
- LLM-Auction: Generative Auction towards LLM-Native Advertising (NeurIPS 2026 · similarity 0.86)
- No-Regret Online Autobidding Algorithms in First-price Auctions (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.