TL;DR
This paper proposes "autobidding" strategies for advertisers to achieve proportional representation of demographic groups in ad campaigns without altering the underlying auction platform. They demonstrate that enforcing fairness on bids and outcomes can drastically reduce advertising efficiency, and advocate for focusing solely on outcome parity. Their autobidder achieves this balance, offering high utility and fair representation.
What Problem Does This Paper Address?
Many companies use online advertising platforms (like Google, Facebook) to recruit for job openings. The goal is often to reach a diverse pool of applicants. However, even if advertisers bid "fairly" (e.g., the same bid for all individuals), the actual outcomes (who sees the ads) might not be proportional across demographic groups. This is because different groups can face varying levels of competition on the platform, making them more or less expensive to reach. Previous work has suggested changing the auction platform itself to fix this, but that's often impractical. This paper tackles the problem from the advertiser's perspective: how can an advertiser achieve fair representation without changing the platform, by adjusting their bidding strategy?
Methods
The core of the paper is developing and analyzing "autobidding" strategies. Instead of manually setting bids, an autobidder algorithm dynamically adjusts bids to meet specific goals. The authors compare three fairness notions:
- Bid Parity: Bidding the same amount for all individuals, regardless of group.
- Outcome Parity: Ensuring the proportion of ad recipients (clicks or impressions) from different groups matches their representation in the population. Bids themselves are not constrained.
- Bid-and-Outcome Parity: A stricter version requiring both bids and outcomes to be proportional.
They formulate the problem as a linear program and develop algorithms to find optimal bidding strategies. Key techniques include:
- Linear Programming Formulation: Representing the problem of maximizing utility subject to budget and representation constraints.
- Randomized Rounding: A technique to convert fractional solutions from the linear program into integer (0 or 1) decisions (bid or don't bid), ensuring constraints are met with high probability.
- Deterministic Rounding: A method for specific cases (like disjoint groups) that guarantees constraints are met deterministically.
- Multiplicative Weights Algorithm: Used to efficiently solve the underlying optimization problems.
Datasets
The authors use data adapted from the American Community Survey (ACS) and the US Bureau of Labor Statistics. They modify this data to simulate real-world scenarios, including varying costs for reaching different demographic groups (specifically, higher costs for women) and estimating click-through rates based on occupation similarity.
Key Insights and Ideas
- Bid Parity is Insufficient: Simply bidding the same amount for everyone doesn't guarantee proportional representation due to varying competition levels across groups.
- Bid-and-Outcome Parity is Inefficient: Requiring both bids and outcomes to be fair can lead to a drastic reduction in advertising utility (efficiency) compared to just focusing on outcome parity. This is because achieving fair bids might force advertisers to overpay significantly for certain groups.
- Autobidding is the Solution: Shifting focus to designing smart bidding strategies (autobidders) that optimize for outcome parity is a more practical and efficient approach.
- Flexibility of Autobidders: Autobidders can be tailored to specific advertiser goals (like proportional representation) and can adapt to market dynamics without requiring platform changes.
- Randomness in Outcomes: Even with deterministic bids, the actual clicks are probabilistic. This necessitates techniques like randomized rounding to provide guarantees on fairness with high probability.
Key Results
- Efficiency Trade-offs: The paper demonstrates with examples that strategies enforcing bid-and-outcome parity can lead to arbitrarily large decreases in advertiser utility compared to strategies enforcing only outcome parity.
- Autobidder Performance: Their proposed autobidder, which focuses on outcome parity, achieves a good balance between high advertiser utility and proportional representation across demographic groups.
- Empirical Validation: Experiments using modified real-world data show that their autobidder outperforms simple bid-parity strategies in achieving representation and often surpasses bid-and-outcome-parity strategies in utility. For instance, in entertainment occupations where women are more expensive to reach, bid parity leads to very low representation, while their autobidder achieves proportional representation with higher utility.
Limitations and Future Work
- Offline Setting: The primary model assumes an "offline" setting where the autobidder has full knowledge of all potential ad slots, costs, and click-through rates. While they extend to an "online" setting, further research is needed for more complex online scenarios.
- Single Advertiser Focus: The current work focuses on a single advertiser. Understanding market dynamics when multiple autobidders with fairness constraints interact is a crucial next step.
- Data Imperfections: Real-world data often has missing features and biases. The authors acknowledge that designing fair mechanisms in the presence of such data challenges is an important future direction.
- Multiple Recruitment Sources: Job recruitment involves various channels (referrals, job boards, etc.), each with different costs and diversity profiles. Exploring composition effects across these sources is a promising area for future research.
- Defining "Fairness": The paper highlights that there are many definitions of fairness, and achieving one (like bid parity) might contradict another (like outcome parity). Further work could explore how to navigate these trade-offs and define "fair" relaxations of constraints.
Summary updated