Platform-supported Auditing Of Social Media Algorithms For Public Interest

on YouTube ↗32m

" summary

Researchers become an advertiser and run paired campaigns (e.g. lumber vs cleaning, Domino’s vs Instacart, Netflix vs NVIDIA) to control for usage, cost and qualifications, finding Facebook’s ad delivery reproduces real-world gender/race skews while LinkedIn does not and that political ads are steered toward like-minded users and cost more to reach dissenters. Proposal builds a platform-supported system that returns noisy relevance scores (via differential privacy) so auditors can test fairness metrics without exposing raw data or trade secrets.

AI-written summary · about

" chapters (27)

1. 0:00Intro
2. 1:24HARMS OF SOCIAL MEDIA ALGORITHMS?
3. 2:09WHAT WE CAN STUDY: ALGORITHMS OF AD DELIVERY
4. 3:23PLAN FOR TODAY
5. 4:02TWO-PHASE AD PROCESS
6. 4:57AD DELIVERY ALGORITHMS BALANCE MANY OBJECTIVES Facebook's business interests
7. 6:16OPAQUE AD DELIVERY ALGORITHMS
8. 7:19IS JOB AD DELIVERY DISCRIMINATORY?
9. 10:37TOOLS FOR BLACK BOX AUDITING
10. 11:372 SIMULTANEOUS ADS, TARGETING A GENDER-AND RACE-BALANCED
11. 14:20APPROACH FOR CONTROLLING FOR QUALIFICATIONS Find a pair of job positions with
12. 15:27DISCRIMINATION IN JOB ADS DELIVERY DRIVER Instacart has de facto more women drivers than Dominas
13. 15:57SIMILAR RESULTS FOR SWE JO
14. 17:02IS POLITICAL AD DELIVERY "ECHO-CHAMBER" LIKE?
15. 17:29FACEBOOK STEERS AD CONTEN TOWARDS "ALIGNED" USERS
16. 17:48MORE EXPENSIVE TO REACH USERS WHO DON'T SHARE CAMPAIGN'S VIEW
17. 19:15PROPOSALS FOR MANDATING PLATFORM SUPPORT • An active area of legislation • Legal framework for vetted researchers to conduct regular audits
18. 20:30CAN / SHOULD PLATFORMS OFFER MORE TRANSPARENCY & SUPPORT AUDITING?
19. 23:30CORE OF THE PROBLEM: "RELEVANCE" ALGORITHMS • Use of algorithms to decide what is relevant to a user • Determine how both organic and promoted content is delivered
20. 24:30UPLOAD CONTENT AND AUDIENCE • The auditor selects and uploads
21. 25:18PLATFORM CALCULATE RELEVANCE SCORES
22. 25:32PLATFORM ADDS NOISE TO PROTECT PRIVACY Why add privacy protection to relevance scores
23. 26:39AUDITOR APPLIES FAIRNESS METRIC
24. 27:03A PLATFORM-SUPPORTED AUDITING FRAMEWORK OF RELEVANCE ALGORITHMS
25. 27:35USE-CASE: BIAS IN JOB AD DELIVERY
26. 27:48SMALL COST OF ENSURING PRIVAC IN JOB AD DELIVERY AUDITING
27. 28:37PLATFORM-SUPPORTED AUDITING IMPLICATIONS

" related talks

• 01h 12mInvestigating algorithmic systems - Algorithm auditing.Marc Rotenberg, Joris van Hoboken, Michael Veale, Bettina Berendt, Juhi Kulshrestha · Privacy Camp 2018
• 28mAdvertising, algorithms and Privacy: recent opensource developmentsRuben Mak · PyData Eindhoven 2019
• 19mMachine Learning for Privacy-Preserving AdsEmanuele Mercanti · DevWorld 2024
• 22mCommunicating Algorithmic Process in Online Behavioral AdvertisingMotahhare Eslami, Sneha R. Krishna Kumaran, Christian Sandvig, Karrie G. Karahalios · CHI 2018
• 01h 24mfrom “old” filter bubbles to political microtargeting.Rocco Bellanova, Frederike Kaltheuner, Katarzyna Szymielewicz, Jeffrey Chester · Privacy Camp 2018
• 12mWhy am I seeing this: Democratizing End User Auditing for Online Content RecommendationsYaxing Yao, Tianshi Li, Leyang Li, Toby Jia-Jun Li, Chaoran Chen, Luke Cao, Yanfang Ye · UIST 2025
• 29mExamining the Relevance of Political Advertising TransparencyNayantara Ranganathan · One Vote Annual Conference 2022
• 54mKeynote - Fairness and Diversity in Online Social SystemsElisa Celis · PyData Berlin 2018