Simultaneous Optimization of Bid Shading and Internal Auction for Demand-Side Platforms

Yadong Xu, Bonan Ni, Weiran Shen, Xun Wang, Zichen Wang, Yinsong Xue, Pingzhong TangAAAI 2024 · DOI

Introduction

In the evolving landscape of online advertising, Demand-Side Platforms (DSPs) face a dual challenge. Historically, ad exchanges operated under second-price auction rules, where truthful bidding was the dominant strategy. However, the industry-wide shift toward first-price auctions—where the winner pays their exact bid—has fundamentally changed the game.

This shift necessitates two critical operations for a DSP: bid shading (lowering bids to capture surplus) and managing internal competition among multiple advertisers vying for the same slot. Traditionally, these have been treated as independent problems, often leading to suboptimal performance due to an "objective mismatch" between the bid landscape estimation and the final auction outcome.

The DSP system model for joint bid shading and internal auction optimization.

The MINMAX Framework

This paper introduces a unified framework that simultaneously optimizes bid shading and internal auction design. The authors provide a theoretical proof that this complex joint optimization can be reduced to learning a set of monotone bid functions for each advertiser. This reduction ensures that the resulting internal auction remains Incentive Compatible (IC), meaning advertisers are naturally incentivized to report their true values.

To implement this, the researchers developed MINMAX, a specialized neural network architecture. Key features include:

  • Partial Monotonicity: The network uses min-max operators over linear functions to guarantee that as an advertiser's value increases, their submitted bid never decreases.
  • End-to-End Learning: Instead of a two-stage process that first models the winning probability distribution, MINMAX uses a surrogate loss function to optimize the final objective (surplus or profit) directly from raw auction data.
  • Invertibility: The architecture allows for the calculation of payments owed by advertisers, a necessary component for a functioning internal auction.

Experimental Results

The framework was evaluated against synthetic data and the industrial iPinYou RTB dataset. A key strength of the MINMAX approach is its robustness across varying bid landscapes. While traditional distribution-based methods (like Dist-L) often fail when the underlying data distribution shifts, MINMAX remains consistently near-optimal.

Learned bid functions for surplus (λ=1) and profit (λ=0) objectives across uniform and exponential distributions.

On industrial benchmarks, MINMAX achieved a 7.42% lift in surplus and a 22.75% lift in profit compared to state-of-the-art baselines. The model's efficiency is also noteworthy; by learning the bid function directly, it avoids expensive numerical searches during live auctions, meeting the millisecond-latency requirements of real-time bidding (RTB).

Real-World Impact and Future Work

Beyond simulations, the model was deployed in a production environment of a global advertising platform. Online A/B testing showed a 6.41% improvement in surplus, demonstrating that the algorithm bids "smarter" by identifying high-value opportunities more effectively rather than simply bidding higher.

Future research directions include adapting the framework for censored auctions, where feedback is only available for winning bids, and integrating long-term budget constraints for advertisers into the joint optimization process.

Summary updated

BibTeX
@inproceedings{xu2024simultaneous,
  author = {Yadong Xu and Bonan Ni and Weiran Shen and Xun Wang and Zichen Wang and Yinsong Xue and Pingzhong Tang},
  title = {{Simultaneous Optimization of Bid Shading and Internal Auction for Demand-Side Platforms}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i9.28855},
}
Figure