Mentors

Kenny Joseph and Atri.

Background

As we have seen/will see in class, output of algorithms might have bias in their output (see e.g. the introductory lecture notes). But many such algorithms are "black-box" in that one does not have access to the internal workings of the algorithm (other than access to input-output behavior). In such a situation is there any hope to figure out if the algorithm is biased?

Algorithmic auditing is a field of research that seeks to assess the biases held by a particular personalization algorithm. It takes ideas that originated in the study of biases in responses to resumes sent in for job searches by Bertrand and Mullainathan and applies them to the web.

Algorithms on a long list of websites have been audited, including (but not limited to):

• Twitter (see e.g. Kulshrestha et al.)
• Youtube (see e.g. Jiang et al.)
• Google Search (see e.g. Robertson et al. (CSCW), Robertson et al. (WWW) and Hannák et al. (2013))
• TaskRabbit (see e.g. Hannák et al. (2017))
• AirBnB (see e.g. Abrahao et al.)

For more examples, see Chen et al. and Hannák et al. (2014).

Finally, for a more a more general overview, see Sandvig et al..

Proposed project

Proposed project-specific deliverables

We first list all the steps that would need to be done to take this project to its logical conclusion:

1. Read the relevant literature to better understand algorithmic auditing (the references below should suffice!)
2. Chose a new algorithm and select a form of bias (e.g. gender or racial discrimination, partisan bias) that you wish to analyze
3. Build a tool to carry out searches on that algorithm. Note: You do not need to perform as extensive an experiment as in the prior work (e.g., you don’t necessarily need to spin up a whole bunch of AWS servers all around the globe)
4. Carry out a statistical analysis that determines the extent to which the algorithm is biased

References

• Bruno Abrahao, Paolo Parigi, Alok Gupta, and Karen S. Cook, Reputation offsets trust judgments based on social biases among Airbnb users . In PNAS 2017.
• Marianne Bertrand and Sendhil Mullainathan, Are Emily and Greg More Employable Than Lakisha and Jamal? A Field Experiment on Labor Market Discrimination . In American Economic Review 2004. [NBER working paper ]
• Le Chen, Anikó Hannák, Ruijin Ma and Christo Wilson, Investigating the Impact of Gender on Rank in Resume Search Engines . In CHI 2018. [Author copy of the paper ]
• Anikó Hannák, Piotr Sapieżyński, Arash Molavi Khaki, David Lazer, Alan Mislove and Christo Wilson, Measuring personalization of web search . In WWW 2013. [More details ]
• Anikó Hannák, Gary Soeller, David Lazer, Alan Mislove and Christo Wilson, Measuring Price Discrimination and Steering on E-commerce Web Sites . In ICM 2014. [Author copy of the paper ]
• Anikó Hannák, Claudia Wagner, David Garcia, Alan Mislove, Markus Strohmaier and Christo Wilson, Bias in Online Freelance Marketplaces: Evidence from TaskRabbit and Fiverr . In CSCW 17.
• Shan Jiang, Ronald E. Robertson and Christo Wilson, Bias Misperceived:The Role of Partisanship and Misinformation in YouTube Comment Moderation . In ICWSM 2019.
• Juhi Kulshrestha, Motahhare Eslami, Johnnatan Messias, Muhammad Bilal Zafar, Saptarshi Ghosh, Krishna P. Gummadi and Karrie Karahalios, Quantifying Search Bias: Investigating Sources of Bias for Political Searches in Social Media . In CSCW 2017. [More details ]
• Ronald E. Robertson and Shan Jiang and Kenneth Joseph and Lisa Friedland and David Lazer and Christo Wilson, Auditing Partisan Audience Bias within Google Search . In CSCW 2018. [Author copy of the paper ]
• Ronald E. Robertson, David Lazer and Christo Wilson, Auditing the Personalization and Composition of Politically-Related Search Engine Results Pages . In WWW 2018. [Author copy of the paper ]
• Christian Sandvig, Kevin Hamilton, Karrie Karahalios and Cedric Langbort, Auditing Algorithms: Research Methods for Detecting Discrimination on Internet Platforms . 2014.