Toward Controlling Discrimination in Online Ad Auctions L. Elisa Celis 1 , Anay Mehrotra 2 , Nisheeth K. Vishnoi 1 1 Yale University 2 IIT Kanpur Ad Exchange Platform User Advertisers Poster: Thursday, June 13 th , 6:30PM-9:00PM @ Pacific Ballroom #125
Online Advertising Online advertising is a major source of revenue for many online platforms, contributing $100+ billion in revenue in 2018. Toward Controlling Discrimination in Online Ad Auctions 6:30 - 9:00 PM @ Pacific Ballroom #125
Discrimination in Online Advertising On Facebook (with 52% women) a STEM job ad was shown to 20% more men than women (Lambrecht & Tucker 2018). Also observed across race (Sweeney 2013) and in housing ads (Ali et al. 2019). User Advertisements User Advertisements Toward Controlling Discrimination in Online Ad Auctions 6:30 - 9:00 PM @ Pacific Ballroom #125
Discrimination in Online Advertising On Facebook (with 52% women) a STEM job ad was shown to 20% more men than women (Lambrecht & Tucker 2018). Also observed across race (Sweeney 2013) and in housing ads (Ali et al. 2019). User Advertisements User Advertisements Can we develop a framework to mitigate this kind of discrimination? Toward Controlling Discrimination in Online Ad Auctions 6:30 - 9:00 PM @ Pacific Ballroom #125
Model and Preliminaries • ! advertisers, " types of users. , as input, mechanism ℳ decides an • For type # ∈ " , r eceiving bids ' ( ∈ ℝ *+ allocation .(' ( ) ∈ [0,1] , and a price 5 ' ( ∈ ℝ , . Choosing the mechanism ℳ , is a well studied problem. Toward Controlling Discrimination in Online Ad Auctions 6:30 - 9:00 PM @ Pacific Ballroom #125
Fairness Constraints Coverage 6 7( : Probability advertiser 8 wins and user Fairness Metric: Equal Representation Constraints: ℓ 7( = ⁄ F G and C 7( = ⁄ F G is of type # For all 8 ∈ ! , # ∈ " ; <= ; <? ≤ C 7( . ℓ 7( ≤ B ∑ ?@A Allows for • constraints on some or all advertisers, • across some or all sub-populations , and • varyi va rying ng the the f fairne rness m metri tric by varying the constraints.. Works for a wide class of fairness metrics; e.g., (Celis, Huang, Keswani and Vishnoi 2019). Toward Controlling Discrimination in Online Ad Auctions 6:30 - 9:00 PM @ Pacific Ballroom #125
Infinite Dimensional Fair Advertising Problem • For many platforms ℳ is the 2 nd price auction. Input: ℓ, C ∈ ℝ H×J Input Output: Set of allocation rules . 7( : ℝ , → 0,1 , • Myerson’s mechanism is the 2 nd price auction on Ou virtual values, P <= ⋅ *+ rev ℳ (. F , . R , … , . J ) max (1) [ ' ≔ ' ⋅ 1 − cdf ' ⁄ pdf ' . J 6 7X . X 6 7( . ( ≥ ℓ 7( ∑ XYF T. V. , ∀ 8 ∈ ! , # ∈ " J 6 7X . X • Let ^ 7( density function of [ 7( (') of advertiser 8 6 7( . ( ≤ C 7( ∑ XYF ∀ 8 ∈ ! , # ∈ " for type # , and _ be the dist. of types. , ∑ 7YF . 7( [ ( ≤ 1 ∀# ∈ " , [ ( . 7( are functions – infinite dimensional optimization problem. • How can we find the optimal . 7( ? Toward Controlling Discrimination in Online Ad Auctions 6:30 - 9:00 PM @ Pacific Ballroom #125
Characterization Result As Assume: • Bids are drawn from a regular distribution. (Equivalent to Myerson.) Toward Controlling Discrimination in Online Ad Auctions 6:30 - 9:00 PM @ Pacific Ballroom #125
Characterization Result As Assume: • Bids are drawn from a regular distribution. (Equivalent to Myerson.) The Then: n: l) There is a “shift” ` ∈ ℝ ,×J such that The Theorem m 4.1 (Inf Informa mal) . 7( ' ( , ` ( ≔ a[8 ∈ argmax ℓ∈[,] ( [ ℓb ' ℓ( + ` ℓ( )] is optimal. Toward Controlling Discrimination in Online Ad Auctions 6:30 - 9:00 PM @ Pacific Ballroom #125
Characterization Result As Assume: • Bids are drawn from a regular distribution. (Equivalent to Myerson.) The Then: n: l) There is a “shift” ` ∈ ℝ ,×J such that Theorem The m 4.1 (Inf Informa mal) . 7( ' ( , ` ( ≔ a[8 ∈ argmax ℓ∈[,] ( [ ℓb ' ℓ( + ` ℓ( )] is optimal. Infinite Dimensional Optimization → Finite Dimensional Optimization. Toward Controlling Discrimination in Online Ad Auctions 6:30 - 9:00 PM @ Pacific Ballroom #125
Algorithmic Result Assume: As ∀ 8 ∈ [!], # ∈ ["] 6 7( > e (Minimum coverage) ∀ ' ∈ supp ^ f J7, ≤ ^ 7( ' ≤ f JgP (Distributed Dist.) 7( ∀ ' F , ' R ∈ supp (^ ^ 7( ' F − ^ 7( (' R ) ≤ h ' F − ' R 7( ) (Lipschitz Cont. Dist.) ∀ 8 ∈ [!], # ∈ ["] i [ 7( ≤ j (Bounded bid) The Then: n: The Theorem m 4.3 (Inf Informa mal) l) There is an algorithm which solves (1) in l ! m n oR log " ⋅ p Bqr s t k (p B<u v) w (h + ! R f JgP R ) steps. Toward Controlling Discrimination in Online Ad Auctions 6:30 - 9:00 PM @ Pacific Ballroom #125
Empirical Results Yahoo! A1 dataset; contains real bids from Yahoo! Online Auctions. Keyword ↔ User type, consider “similar” keywords pairs. : " = 2, C 7( = 1, and #auctions = 3282 . Se Settin ing: Vary: ℓ 7( = ℓ ∈ 0,0.5 Va Me Measures: ⁄ Fairness slift ℱ ≔ min 7( 6 7( (1 − 6 7( ) , and ⁄ Revenue ratio É ℳ,ℱ ≔ rev ℳ rev ℱ . ( ) Toward Controlling Discrimination in Online Ad Auctions 6:30 - 9:00 PM @ Pacific Ballroom #125
Conclusion and Future Work We give an optimal truthful mechanism which pr bly satisfies fairness provably constraints and an efficient algorithm to find it. We observe a minor loss to the revenue and change to advertiser distribution when using it. • How does the mechanism affect user and advertiser satisfaction? • Can we incorporate asynchronous campaigns? • Can we extend our results to the GSP auctions? Tha Thank nks! s! Poster: Thursday, June 13 th , 6:30PM-9:00PM @ Pacific Ballroom #125 Toward Controlling Discrimination in Online Ad Auctions 6:30 - 9:00 PM @ Pacific Ballroom #125
Recommend
More recommend
