Abstract
Online advertising channels have commonly focused on maximizing total advertiser value (or welfare) to enhance long-run retention and channel healthiness. Previous literature has studied auction design by incorporating machine learning predictions on advertiser values (also known as machine-learned advice) through various forms to improve total welfare. Yet, such improvements could come at the cost of individual bidders' welfare and do not shed light on how particular advertiser bidding strategies impact welfare. Motivated by this, we present an analysis on an individual bidder's welfare loss in the autobidding world for auctions with and without machine-learned advice, and also uncover how advertiser strategies relate to such losses. In particular, we demonstrate how ad platforms can utilize ML advice to improve welfare guarantee on the aggregate and individual bidder level by setting ML advice as personalized reserve prices when the platform consists of autobidders who maximize value while respecting a return-on-ad spent (ROAS) constraint. Under parallel VCG auctions with such ML advice-based reserves, we present a worst-case welfare lower-bound guarantee for an individual autobidder, and show that the lower-bound guarantee is positively correlated with ML advice quality as well the scale of bids induced by the autobidder's bidding strategies. Further, we prove an impossibility result showing that no truthful, and possibly randomized mechanism with anonymous allocations can achieve universally better individual welfare guarantees than VCG, in presence of personalized reserves based on ML-advice of equal quality. Moreover, we extend our individual welfare guarantee results to generalized first price (GFP) and generalized second price (GSP) auctions.
Introduction
In the modern "autobidding" landscape, advertisers have transitioned from manual bidding to using algorithms that maximize total value (e.g., clicks or conversions) subject to a Return-on-Ad-Spend (ROAS) constraint. While traditional auction design focuses on maximizing "total welfare" for the platform's overall efficiency, this aggregate focus often neglects the welfare of individual bidders.
A significant issue in these environments is the "Bully Problem": a large advertiser with high profit margins in one auction can afford to overbid in another, effectively squeezing out smaller, more efficient competitors. This manipulative power threatens the long-term diversity and health of the advertising ecosystem. This paper addresses this by asking: How can a platform guarantee a minimum level of welfare for an individual advertiser, and how can machine learning (ML) help?
Problem Formulation and Notation
To analyze these dynamics, we consider a set of parallel auctions where bidders submit bids across multiple queries. The following table summarizes the notation used to define the bids, outcomes, and payments within this framework.
The core challenge is to protect the individual welfare of a target bidder $i$ against the worst-case bidding strategies of their competitors. We introduce a strategy-dependent individual welfare metric that compares the worst-case welfare an advertiser receives under any feasible bidding environment to the welfare they would achieve in a perfectly efficient "global welfare maximizing" outcome.
Methodology: ML-Advice and Personalized Reserves
The proposed framework utilizes Machine-Learned (ML) advice—specifically, a lower-confidence bound on an advertiser's true value—to set personalized reserve prices. These reserves act as a "floor," making predatory overbidding more expensive for competitors and safeguarding the "target" bidder.
Mechanism Design
The analysis primarily focuses on Parallel VCG (Vickrey-Clarke-Groves) auctions. We prove that by setting reserves based on $\beta$-accurate ML advice (where $\beta \in [0, 1]$ represents prediction quality), the platform can significantly limit the manipulative power of competitors.
Key theoretical contributions include:
- Welfare Lower Bound: A proof of a worst-case lower bound for individual welfare under VCG with ML-based reserves.
- Impossibility Result: A proof that no truthful, anonymous auction mechanism can provide a strictly better individual guarantee than VCG when using ML-based reserves of equal quality.
- Extensions: Generalization of these findings to Generalized Second Price (GSP) and Generalized First Price (GFP) auctions, identifying conditions where these formats might offer competitive individual guarantees.
Experimental Results
The theoretical findings were validated using semi-synthetic data derived from actual ad auction logs of a major search platform, including slot counts, click-through rates (CTRs), and conversion values.
Key Findings
- Correlation with ML Quality: There is a direct, positive correlation between the accuracy of ML predictions ($\beta$) and the strength of the welfare guarantee. Better AI leads to a fairer auction for the individual.
- Numerical Confirmation: Simulations showed that as ML accuracy improved, the distribution of advertiser welfare shifted toward the efficient ideal. For instance, with no reserves ($\beta=0$), 17% of bidders lost at least 20% of their potential welfare. With high-quality ML advice ($\beta=0.75$), this figure dropped to 8%.
- Bidding Strategy Impact: The welfare guarantee is also positively correlated with the scale of bids induced by the autobidder's own strategy ($\alpha$).
Conclusion
This work shifts the focus of auction design toward individual fairness in the autobidding era. By leveraging ML advice to set personalized reserves, platforms can provide robust welfare guarantees for individual bidders. While the "tighter" welfare bounds proposed are computationally intensive and rely on the assumption of ML advice as a true lower bound, this research provides a foundational theoretical framework for protecting advertiser diversity in complex, algorithmic markets. Future work will explore these guarantees in dynamic, multi-round environments where bidders adapt over time.
Summary updated