Abstract
Digital advertising platforms often produce skewed demographic outcomes in job recruitment due to varying market competition across user groups. This paper presents an "autobidder" framework that enables advertisers to achieve proportional representation (Outcome Parity) without requiring changes to the underlying auction mechanisms of major platforms. By formulating the bidding process as a constrained optimization problem and utilizing Multiplicative Weights Update (MWU) and rounding techniques, the authors demonstrate that advertisers can maintain high utility while ensuring diverse audience reach. The framework is evaluated using US census data, showing that it outperforms traditional bidding strategies in both fairness and efficiency.
Introduction
When organizations utilize platforms like Facebook or Google for recruitment, they frequently aim to attract a diverse and representative pool of applicants. However, a significant structural hurdle exists: market competition is not uniform across demographic groups. Previous research indicates that it is often more expensive to display ads to women than to men because women are more heavily targeted by high-bidding retail brands.
If a recruiter employs a standard "Bid Parity" strategy—bidding the same amount for every user regardless of their demographic—the platform's auction mechanism will naturally favor the "cheaper" group. This results in a skewed audience that fails to meet diversity goals. While some have suggested redesigning auction systems at the platform level, such changes are complex and unlikely to be implemented by major providers. This paper addresses the problem from the advertiser's perspective, proposing a client-side bidding strategy that guarantees fair outcomes even within existing, unchanged auction environments.
The Framework: From Bid Parity to Outcome Parity
The core contribution of this research is a shift in focus from "equal bids" to "equal outcomes." The authors argue that fairness in recruitment advertising should be defined by Outcome Parity—ensuring the final set of individuals who see or click an ad is representative of the target population—rather than bidding the same amount for every impression.
Methodology
The authors frame the challenge as an optimization problem, utilizing several key technical components:
- Linear Programming (LP): The auction is modeled as an integer linear program where the objective is to maximize total value (such as clicks or impressions) subject to budget constraints and specific group representation requirements.
- Multiplicative Weights Update (MWU): To solve the LP efficiently, the authors adapt an algorithmic framework that allows the autobidder to find optimal bidding thresholds for different demographic groups dynamically.
- Rounding Techniques: Because auctions are discrete (an advertiser either wins or loses), the authors developed randomized and deterministic rounding methods. These ensure that theoretical winning probabilities are translated into real-world bids that satisfy fairness guarantees with high probability.
- Online Stochastic Optimization: The model is extended to an online setting, allowing the advertiser to "learn" the optimal bidding strategy over time. This approach achieves "sublinear regret," meaning the strategy's performance improves as the campaign progresses.
Experimental Evaluation
The researchers evaluated their algorithms using data adapted from the 2021 American Community Survey (ACS) for individuals in California and the US Bureau of Labor Statistics. These datasets provided realistic gender and race distributions across various job sectors, such as "Entertainment" and "Computer/Mathematical" fields.
To simulate real-world competitive dynamics, the researchers applied a 10% "price bump" to the cost-per-click (CPC) for women, reflecting the higher competition from retail advertisers observed in empirical studies.
Key Insights and Results
The Inefficiency of Bid-and-Outcome Parity
A striking finding of the study is the inherent inefficiency of requiring both bid parity and outcome parity. The paper proves that forcing an advertiser to bid equally while demanding a fair outcome can lead to an arbitrarily large decrease in utility. Conversely, allowing bids to vary across groups to "buy" the necessary representation is significantly more efficient.
Performance and Utility Gains
In simulations, the proposed autobidder achieved proportional representation while maintaining significantly higher utility (more total clicks/views) than strategies attempting to enforce both bid and outcome parity. The autobidder consistently outperformed:
- Simple Bid Parity: Which failed to reach the more expensive demographic (women).
- Bid-and-Outcome Parity: Which proved too expensive and reduced the overall volume of the campaign.
Practical Feasibility
Unlike rigid quota systems that might be impossible to meet under tight budgets, the authors' ratio-based constraints are technically feasible. If a specific goal is unattainable given the budget and market conditions, the algorithm can notify the advertiser, providing transparency that current bidding tools lack.
Conclusion and Future Directions
By implementing fairness at the bidding level, advertisers can take immediate action toward equitable recruitment without waiting for platform-level changes. However, the authors note several areas for future research:
- Market Equilibrium: The study focuses on a single advertiser. If many advertisers adopt fairness-aware autobidders, it could shift overall market prices (CPC), requiring a study of market equilibrium.
- Information Accuracy: The model assumes the advertiser has accurate estimates of user demographics and costs, which are often noisy in practice.
- The Recruitment Pipeline: Advertising is only the first step. Future work should explore how these bidding strategies interact with subsequent stages of hiring, such as resume screening and interviewing, to ensure fairness throughout the entire funnel.
Summary updated