direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments

Algorithmic Fairness in Rankings of People

People search engines are increasingly common for job recruiting, for finding a freelancer, and even for finding companionship or friendship. As in other cases, a top-k ranking algorithm is used to find the most suitable way of ordering the items (persons, in this case), considering that if the number of people matching a query is large, most users will not scan the entire list. Conventionally, these lists are ranked in descending order of some measure of the relative quality of items (e.g. years of experience or education, upvotes, or inferred attractiveness). Unsurprisingly, the results of these algorithms potentially have an impact on the people who are ranked, and contribute to shaping the experience of everybody online and offline. It is therefore of societal and ethical importance to ask whether these algorithms eventually produce results that demote, marginalize, or exclude individuals belonging to an unprivileged group or a minority. These algorithms may have discriminatory effects, even in the absence of discriminatory intent, imposing a less favorable treatment to already disadvantaged groups. These problems are exacerbated when details about the algorithms used for ranking are unknown.

Given is the definition and solution of the Fair Top-k Ranking problem, in which we want to determine a subset of k candidates from a large pool of n > k candidates, such that we select the “best” candidates
subject to group and individual fairness criteria. Our ranked group fairness definition extends group fairness
using the standard notion of protected group, (e.g. “people with disabilities”) and is based on ensuring that the proportion of protected candidates in every prefix of the top-k ranking is statistically indistinguishable from a target proportion. The ranked group fairness criterion is limited to the case of a binary problem (a candidate can have at most one protected attribute that is not multinomial). This work shall extend the framework to the multinomial and combinatorical case.

Requirements:

basic knowledge in statistics, basic knowledge in Python or a comparable script language, basic knowledge in Git or a comparable version control system

Start:

immediately

Contact:

meike.zehlike@tu-berlin.de

Zusatzinformationen / Extras

Quick Access:

Schnellnavigation zur Seite über Nummerneingabe

Auxiliary Functions

Contact

Meike Zehlike
+49 (30) 314-27998
Room E-N 173