Optimal Auction Design in the Joint Advertising¶
Conference: ICML2025
arXiv: 2507.07418
Code: Not released
Area: Auction Design / Automated Mechanism Design
Keywords: Joint Advertising, Optimal Auction, BundleNet, Incentive Compatibility, Neural Network Mechanism Design
TL;DR¶
This paper proposes an optimal auction mechanism for joint advertising scenarios (where retailers and suppliers co-bid for ad slots): for a single slot, a Myerson-style closed-form optimal solution is derived; for multiple slots, a BundleNet neural network is designed to construct IC constraints on a per-bundle basis, maximizing platform revenue while ensuring approximate incentive compatibility.
Background & Motivation¶
- Joint Advertising is an emerging online advertising paradigm (already deployed on platforms like Facebook): while traditional advertising involves only unilateral bidding by retailers, joint advertising allows retailers and suppliers (or brands) to co-bid for an ad slot, with the platform charging both parties.
- Existing methods suffer from two major limitations:
- VCG/JAMA-like methods establish incentive compatibility (IC) only along a single dimension (either brand or retailer), ignoring the bundle structure, which leaves room for optimization in allocation efficiency.
- JRegNet (a neural-network-based method built on RegretNet) exhibits poor generalization, with performance collapsing even below VCG in large-scale, complex bipartite graph scenarios.
- Key Challenge: In joint advertising, a bundle is co-determined by the strategies of two separate bidders. Because the bundle itself lacks an independent bidding strategy, directly defining bundle-level IC constraints is highly challenging.
Method¶
Problem Modeling¶
In a joint advertising system, the set of retailers \(R\) and the set of suppliers \(S\) are disjoint, and the set of advertisements \(E \subseteq R \times S\) forms a bipartite graph \(G=(R,S,E)\). Each advertisement \(e=(r,s)\) links a retailer and a supplier. The click-through rates (CTRs) for \(m\) slots are denoted by \(\boldsymbol{\lambda}=(\lambda_1,\dots,\lambda_m)\), satisfying \(\lambda_1 \geq \cdots \geq \lambda_m\).
An auction mechanism \(\mathcal{M}=(x,p)\) consists of allocation rules and payment rules:
It must satisfy both Incentive Compatibility (IC) (truthful bidding is a dominant strategy) and Individual Rationality (IR) (non-negative utility for participation).
Optimal Single-Slot Mechanism (Theorem 4.3)¶
Under regular distribution assumptions, the virtual value function is defined as:
The virtual value of bundle \(e=(r,s)\) is: \(c^e(v_r, v_s) = c_r(v_r) + c_s(v_s)\)
The optimal mechanism takes the form of a step function: allocation occurs when the bid exceeds a critical value \(\hat{v}_i(v_{-i})\), and the payment is exactly equal to this critical value. The critical threshold is determined such that the virtual value of the optimal bundle containing bidder \(i\) must simultaneously satisfy: (1) being no less than the platform reserve price \(v_0\); and (2) being no less than the virtual value of any bundle that does not contain \(i\).
Multi-Slot BundleNet¶
Bundle-level IC Constraints (Core Innovation)¶
Unlike JRegNet which constructs separate IC constraints for each bidder, this work proposes a bundle-based IC constraint:
Lemma 5.1 guarantees that \(\sum_{i \in R \cup S} rgt_i(w) \leq \sum_{e \in E} rgt^e(w)\), which means individual IC constraints approach zero as the bundle-level constraints approach zero.
Network Architecture¶
- Graph Feature Fusion: Aggregates bipartite node features into edge features
- Divided Bids: \(DB^e = [X_r, X_s] \in \mathbb{R}^{2m}\) (concatenation)
- Stacked Bids: \(SB^e = X_r + X_s \in \mathbb{R}^m\) (summation)
- Allocation Network: MLP processes \(SB^E\) \(\to\) doubly stochastic matrix approach (taking the minimum of row/column softmax) to ensure allocation feasibility.
- Payment Network: MLP processes \(DB^E\) \(\to\) Sigmoid structures normalize payment coefficients \(\tilde{p} \in [0,1]\), multiplied by the allocation-weighted bid to obtain the actual payment, structurally guaranteeing IR.
Optimization and Training¶
Augmented Lagrangian method is adopted to solve the constrained optimization:
Network parameters \(w\), misreported bids \(v'\), and Lagrange multipliers \(\mu\) are updated alternately.
Key Experimental Results¶
| Scenario | Method | U2 | U3 | U4 | U5 | E2 | E3 | E4 | E5 |
|---|---|---|---|---|---|---|---|---|---|
| Single-slot | BundleNet | 0.529 | 0.668 | 0.781 | 0.880 | 0.425 | 0.546 | 0.635 | 0.722 |
| Single-slot | Optimal | 0.525 | 0.671 | 0.783 | 0.882 | 0.425 | 0.548 | 0.647 | 0.738 |
| Single-slot | JRegNet | 0.562 | 0.729 | 0.779 | 0.788 | 0.473 | 0.589 | 0.631 | 0.694 |
| Single-slot | RVCG | 0.381 | 0.600 | 0.746 | 0.861 | 0.282 | 0.465 | 0.593 | 0.704 |
| Scenario | Method | U5×5 | U8×5 | U10×5 | U12×5 |
|---|---|---|---|---|---|
| 5-slot | BundleNet | 1.498 | 2.089 | 2.405 | 2.650 |
| 5-slot | JRegNet | 1.497 | 1.935 | 1.962 | 1.997 |
| 5-slot | RVCG | 0.842 | 1.883 | 2.280 | 2.587 |
- In single-slot settings, BundleNet almost perfectly approximates the theoretical optimal solution. Although JRegNet sometimes yields higher revenue, it violates the optimal mechanism (cheats on IC constraints).
- In multi-slot settings, BundleNet-5 outperforms RVCG and JRegNet across the board, with the advantage becoming more pronounced as the number of bundles increases.
- All methods maintain an IC violation of < 0.001.
Highlights & Insights¶
- Solid Theoretical Contribution: It derives the necessary and sufficient conditions for the optimal single-slot auction under joint advertising for the first time (Theorem 4.3), successfully extending Myerson's lemma to bundle contexts.
- Elegant Bundle-level IC Constraints: Instead of directly defining strategies for bundles (which is impossible since they lack independent strategies), the paper proves via Lemma 5.1 that bundle-level constraints upper-bound individual-level constraints, resolving the core challenge.
- Problem-aligned Network Design: The graph feature fusion leverages the node-edge correspondence in bipartite graphs. Meanwhile, the allocation network employs doubly stochastic matrices to guarantee feasibility, and the payment network uses a Sigmoid structure to structurally ensure IR.
- Closed-loop Experimental Validation: Single-slot experiments validate correctness by comparing with the theoretical optimum, while multi-slot experiments demonstrate practical superiority.
Limitations & Future Work¶
- Regular Distribution Assumption: The single-slot optimal results rely on distribution regularity (monotonically increasing virtual values). Non-regular distributions would require ironing processes.
- Single-parameter Bidders Only: Each bidder has only one private value, which has not been extended to multi-item or multi-parameter settings.
- Randomly Generated Bipartite Graph Structures: The retailer-supplier relationship matrices in the experiments are randomly generated, and have not yet been evaluated end-to-end in real-world large-scale advertising systems.
- Approximate IC Guarantees: Although the empirical IC violation is extremely low, the work lacks a theoretical bound on approximate IC.
- Scalability: Computational overhead and performance scaling under larger scenarios (e.g., hundreds of bundles) are not reported.
Related Work & Insights¶
- Myerson (1981): The classic single-parameter optimal auction theory. This work extends it to the bundle scenarios of joint advertising.
- RegretNet (Dütting et al., 2024): Pioneering work utilizing neural networks to approximate optimal mechanisms. BundleNet improves upon its architecture and constraints.
- JRegNet (Zhang et al., 2024): A RegretNet variant for joint advertising, representing the direct benchmark for this study.
- JAMA (Ma et al., 2024): An AMA-based approach for joint advertising that established the joint advertising setting.
Rating¶
- Novelty: ⭐⭐⭐⭐ (the bundle-level IC constraint and single-slot optimal characterization are highly original)
- Experimental Thoroughness: ⭐⭐⭐⭐ (comprehensive coverage of single/multi-slot settings and various distributions, though lacks real-world deployment evaluation)
- Writing Quality: ⭐⭐⭐⭐ (clear structure and robust connection between theory and experiments)
- Value: ⭐⭐⭐⭐ (joint advertising is a newly deployed practical setting, with dual contributions in both theory and algorithms)