---
title: 'JEANet: Unified AMD for Joint Ad Auctions'
url: https://www.emergentmind.com/topics/jeanet
type: topic
---

# JEANet: Unified AMD for Joint Ad Auctions

JEANet (Joint Auction with Externalities and Adaptation Network) is an automated mechanism design (AMD) framework for multi-slot advertising auctions that accommodates both traditional and joint advertising formats, models global externalities affecting user experience, and adapts to heterogeneous bidding behavior. It produces allocation and payment mechanisms that provably satisfy individual rationality (IR) and approximate dominant strategy incentive compatibility (DSIC), optimizing a tunable combination of platform revenue and user experience. JEANet is distinguished as the first AMD approach to incorporate global externalities—quantifying page-level effects due to ad and organic item placement—and achieve unified auctions integrating separate and bundled (store-brand) advertisers [2512.15043].

## 1. Problem Domain and Motivation

JEANet addresses the allocation of $K$ display slots among heterogeneous contenders: traditional advertisers bidding scalars ($b_i$), joint advertisers bidding vectors for store–brand bundles ($(b_{store}, b_{brand})$), and organic items. The challenge lies in maximizing a composite objective—weighted platform revenue and user experience—while enforcing mechanism-theoretic constraints (IR, approximate DSIC) and handling variability in bidding protocols due to multi-party participation and joint selling relationships. In contrast to prior work, JEANet’s formulation subsumes both traditional and joint slot auctions, explicitly models global externalities via item-specific user experience scores $ue_i$, and handles a mixed inventory of organic and paid candidates.

## 2. Mechanism-Design Formulation

JEANet operationalizes the auction as a mechanism $\mathcal{M} = (\mathbf{a}, \mathbf{p})$ with discrete allocation variables $a_{ik} \in \{0,1\}$ specifying item-slot assignments under feasibility constraints:

\[
\sum_{k=1}^K a_{ik} \leq 1 \;\forall i; \qquad \sum_{i=1}^{m+n} a_{ik} = 1 \;\forall k 
\]

with $m$ advertisers and $n$ organic items. The expected click-through rate (CTR) for item $i$ is $a_i = \sum_{k=1}^K a_{ik} \alpha_k$, with $\alpha_k$ denoting slot-specific CTRs. Payments $p_i$ are computed per item, and each advertiser’s utility is quasi-linear:

\[
u_i(v_i;\mathbf{b}, \mathbf{ue}, \mathbf{c}, \boldsymbol{\alpha}) = v_i a_i(\cdot) - p_i(\cdot)
\]

Individual Rationality is enforced:

\[
u_i(v_i; (v_i, \mathbf{b}_{-i}), \cdots) \geq 0,\;\forall i, v_i
\]

Approximate DSIC is controlled via ex-post regret:

\[
\mathrm{rgt}_i(\mathbf{v}, \mathbf{ue}, \ldots) = \max_{b'_i} \left[ u_i(v_i; (b'_i, \mathbf{b}_{-i}), \ldots) - u_i(v_i; (v_i, \mathbf{b}_{-i}), \ldots) \right]
\]

JEANet’s mechanisms target $\mathrm{E}[\mathrm{rgt}_i] \approx 0$. The principal objective maximizes expected revenue plus a discounted user experience sum, with trade-off hyperparameter $\gamma$:

\[
\max_{(\mathbf{a}, \mathbf{p})} \; \mathbb{E} \left[ \sum_i p_i + \gamma \sum_i ue_i a_i \right]
\]

subject to IR, approximate DSIC, and feasibility.

## 3. Network Architecture

JEANet parameterizes allocation $\mathbf{a}^w$ and payments $\mathbf{p}^w$ via neural network weights, implemented with three interoperable modules:

### 3.1 Adaptive Extraction Module (AEM)

- Contextual embeddings: $t_i = \mathrm{Emb}(c_i) \in \mathbb{R}^{d_c'}$ for each item context.
- Bid profile and user experience embeddings for slot $k$:
  - $e_{ik} = b_i \cdot \alpha_k$ (for ads); $z_{ik} = \mathrm{Emb}(ue_i)\,\alpha_k$.
  - Organic items assigned $e_{jk} = 0$ for $j > m$.
- Residual-Quantization (RQ) layers discretize $\mathbf{e}$ into codebook indices $(u_1, \dots, u_D)$, reconstruct $\widehat{\mathbf{e}}$.
- Concatenation and linear reduction yield $T \in \mathbb{R}^{(m+n) \times K \times d'}$.

### 3.2 Externality Transformer Module (ETM)

- Row-wise transformer per item models interactions among slot choices.
- Column-wise transformer per slot captures competitive dynamics across all items.
- Global average derives page-level features.
- Fused output: $O \in \mathbb{R}^{(m+n) \times K \times d_{out}}$.

### 3.3 Deep Mechanism Module (DMM)

- Five output tensors: $O^r$ (row), $O^c$ (column), $O^a$ (allocation modulator), $O^{p_1}, O^{p_2}$ (payment signals).
- Row-wise and column-wise softmaxes yield $\hat{O}^r$, $\hat{O}^c$ enforcing slot and item constraints.
- Allocation: $\hat{a}_{ik} = \min\{\hat{O}^r_{ik}, \hat{O}^c_{ik}\}$, modulated by $\sigma(O^a_{ik})$ to final $a^w_{ik}$.
- Payments: $\tilde{p}_{i,j} = \sigma(\frac{1}{K} \sum_k O^{p_j}_{ik}) \in (0,1)$; $p^w_{i,j} = \tilde{p}_{i,j} \sum_{k=1}^K a^w_{ik} e_{ik}$.
- Slot assignment finalized through Birkhoff-von Neumann rounding.

## 4. Training Procedure

Training alternates between two phases:

### 4.1 Pretraining AEM

- Loss: $L_{AEM} = ||\mathbf{R} - \widehat{\mathbf{R}}||^2_2 + \beta \sum_{d=1}^D ||\mathbf{e} - \mathrm{sg}[\widehat{\mathbf{e}}^{(d)}]||^2_2$
- Residual-Quantization and straight-through estimators support quantized feature learning.

### 4.2 Joint Training of DMM with Regret Penalty

- Regrets $\widehat{\mathrm{rgt}_i(w)}$ are computed by optimizing worst-case misreports $b'_i$.
- Augmented Lagrangian loss:

\[
C_\rho(w;\lambda) = -\frac{1}{L}\sum_{\ell,i} \left[ p^w_i + \gamma \sum_j ue_j a^w_j \right]_{(\ell)} + \sum_i \lambda_i \widehat{\mathrm{rgt}_i(w)} + \frac{\rho}{2} \sum_i \widehat{\mathrm{rgt}_i(w)}^2
\]

- Training alternates between:
  - Computing per-sample regrets,
  - Gradient descent on network weights $w$,
  - Gradient ascent on multipliers $\lambda_i$.

## 5. Algorithmic Structure

The training and deployment algorithm can be summarized:

```python
# 1. Pretrain AEM
for epoch in 1..N_aem:
    for minibatch B:
        compute e, z → R
        quantize e → ŷ_e → reconstruct Ŕ
        lossA = ||R−Ŕ||² + β·commitment(e, ŷ_e)
        θ ← θ − η_aem · ∇_θ lossA
Freeze θ

# 2. Jointly train DMM with regret penalty
for epoch in 1..N_mech:
    for minibatch B:
        # forward pass to get a^w, p^w
        # compute regret by max misreport
        compute rgt_i for all i
        lossMech = − {Rev + γ·UE} + Σ_i [λ_i·rgt_i + (ρ/2)·rgt_i²]
        w ← w − η_mech · ∇_w lossMech
        λ_i ← λ_i + η_mech · rgt_i

Output: trained JEANet(θ, w)
```

## 6. Experimental Results

Extensive experiments confirm JEANet’s efficacy across synthetic and industrial auction data. Synthetic setups cover random item counts (3–10 ads of each type, 20 total items), slot configurations (3–6 slots), and multiple value distributions (Uniform, Normal, Multi-lognormal). Industrial experiments utilize 15 days of real GSP logs, filtered to consistent $(m, n, K)$ tuples.

### Baseline Mechanisms

- GSP with fixed slots
- VCG (social-welfare maximizer)
- IAS (Myerson-based)
- Revised JRegNet (bundle-based RegretNet)
- Revised TICNet (contextual mechanism)

### Evaluation Metrics

Metrics are normalized such that VCG=1, Score=1.5. Primary metrics include:

| Abbreviation | Quantity              | Formula                          |
|--------------|-----------------------|----------------------------------|
| SW           | Social Welfare        | -                                |
| Rev          | Revenue               | -                                |
| UE           | User Experience       | -                                |
| Score        | Overall Score         | $Rev + 0.5 \cdot UE$             |

JEANet achieves uniformly highest Score ($p < 0.05$) in every tested scenario, improving Score by 10–20% compared to best existing baselines. On industrial data with 4–6 slots, Score is raised from approximately 2.3 to 3.2, maintaining near-optimal UE and elevating Rev. Ablation confirms architectural necessity: excising AEM or replacing ETM yields a 4–8% Score reduction.

A plausible implication is that Transformer-based global modeling and adaptive bid feature extraction are prerequisites for robust high-performing joint auction mechanisms in heterogeneous settings.

## 7. Contextual Significance

JEANet represents the first AMD-based mechanism that fully unifies traditional and joint advertising, directly models global ads–organic externalities, and provably enforces IR plus approximate DSIC in realistic auctions. The framework adapts to variable bidding distributions, supports complex bundle-bidding, and ensures trade-off optimality for revenue and experience objectives across both synthetic and real-marketplace deployments [2512.15043]. This suggests broad applicability to online advertising platforms seeking mechanism-driven improvements in allocation efficiency and platform utility.

Source: https://www.emergentmind.com/topics/jeanet