Fairness Through Awareness

Fairness Through Awareness Cynthia Dwork Moritz Hardt Toniann Pitassi Omer Reingold Richard Zemel

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Fairness Through Awareness 2012 ITCS / arXiv 10.1145/2090236.2090255 1104.3913 dwork2012-fairness dwork2012-fairness.txt, the project's staged copy 3cbe51a7719ef45c5db1f23d04bc817468e0f6a8 arxiv-open law

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them method Quotation withheld because this record is not cleared for redistribution. fairness in classification Quotation withheld because this record is not cleared for redistribution. formal-model formal-model Quotation withheld because this record is not cleared for redistribution. mechanisms fairness as treat similar individuals similarly, formalized as a Lipschitz condition on the classifier w.r.t. a task-specific similarity metric, maximizing utility subject to that constraint Quotation withheld because this record is not cleared for redistribution. the approach assumes a task-specific similarity metric, which is hard to define what this limits: every guarantee the framework offers: it assumes Quotation withheld - supplied from outside the definition Quotation withheld because this record is not cleared for redistribution. Obtaining or justifying the task-specific similarity metric the framework depends on is the hardest part, and any real metric will be an approximation. what this limits: the practical applicability of the whole approach: JUSTIFYING THE AVAILABILITY OF OR ACCESS TO THE DISTANCE METRIC ... IS ONE OF THE MOST CHALLENGING ASPECTS OF OUR FRAMEWORK, AND IN REALITY THE METRIC USED WIL[l be imperfect] Quotation withheld because this record is not cleared for redistribution. The general fairness LP gives no quantitative bound on the loss it incurs. what this limits: the loss claim for the general fairness LP: THE RESULT HOWEVER DOES NOT GIVE A CONCRETE QUANTITATIVE BOUND ON THE RESULTING LOSS Quotation withheld because this record is not cleared for redistribution. When the protected set is far from the rest of the population under the metric, satisfying the Lipschitz condition can still leave unfair practices unchecked. what this limits: the protection guarantee for a distant protected set: IF MEMBERS OF S ARE RELATIVELY FAR FROM THE MEMBERS OF T ACCORDING TO THE DISTANCE METRIC, THEN SATISFYING THE LIPSCHITZ CONDITION MAY FAIL TO PREVENT SOME OF THE UNFAIR PRACTICES Quotation withheld because this record is not cleared for redistribution. No tight Earthmover characterisation is obtained for the relative-infinity distance version of the bias. what this limits: the Earthmover characterisation: for biasD-infinity WE DO NOT GET A TIGHT CHARACTERIZATION Quotation withheld because this record is not cleared for redistribution. The whole treatment assumes one metric applied uniformly to everyone; per-user or plural metrics are untested. what this limits: the scope of every result in the paper: OUR DISCUSSION THROUGHOUT THIS PAPER HAS ASSUMED THAT A SINGLE METRIC IS USED ACROSS THE BOARD - per-user or plural metrics are untested Quotation withheld because this record is not cleared for redistribution. The nearest existing similarity system the authors point to (AALIM, health care) does not supply the kind of distance metric the framework needs. what this limits: the claim that a metric could be obtained in practice: the nearest existing system DOES NOT (YET?) PROVIDE THE DISTANCE METRIC THAT OUR APPROACH REQUIRES Quotation withheld because this record is not cleared for redistribution. A widely used fairness criterion can hold while the outcome is, in the authors' words, blatantly unfair to an individual. what this limits: statistical parity used as evidence that a classifier is fair: the authors give examples in which statistical parity is maintained while the outcome from an individual's point of view is blatantly unfair Quotation withheld because this record is not cleared for redistribution. The metric the formal guarantee consumes is a social agreement that is expected to change. what this limits: any fairness guarantee obtained in this framework: when ground truth is unavailable the metric may reflect the best available approximation as agreed upon by society, and is assumed to be public and open to continual refinement Quotation withheld because this record is not cleared for redistribution. The bridge from individual fairness to group fairness holds only under a stated distance condition. what this limits: the claim that individual fairness delivers group fairness: it holds Quotation withheld Quotation withheld because this record is not cleared for redistribution. The population the metric is defined over is a modelling choice the authors flag. what this limits: the scope of the fairness guarantee proved here: the authors state they have assumed the set is that of real individuals rather than all possible encodings of individuals Quotation withheld because this record is not cleared for redistribution. differential-privacy tools and a Lipschitz metric formalism applied to define and enforce fairness in classification Quotation withheld because this record is not cleared for redistribution. statistical parity statistical parity Quotation withheld because this record is not cleared for redistribution. group fairness group fairness Quotation withheld because this record is not cleared for redistribution. Lipschitz condition Lipschitz condition Quotation withheld because this record is not cleared for redistribution. Earthmover distance Earthmover distance Quotation withheld because this record is not cleared for redistribution. fairness through blindness fairness through blindness Quotation withheld because this record is not cleared for redistribution. fairness - INDIVIDUAL VERSUS GROUP, AND THE SECOND CAN HOLD WHILE THE FIRST FAILS: individual fairness is the principle that two people similar for a task be classified similarly, statistical parity is a property of demographics in aggregate, and the paper's examples keep the second while the first is blatantly violated Quotation withheld because this record is not cleared for redistribution. the similarity metric - AS SOCIETY'S CURRENT BEST APPROXIMATION: where no ground truth exists the metric stands in for one by agreement, is meant to be public and revisable, and is the input on which every formal result in the paper depends Quotation withheld because this record is not cleared for redistribution. fairness as a principle - TREAT SIMILAR INDIVIDUALS SIMILARLY: the paper's whole notion is one sentence, and everything formal in it is machinery for making 'similar' precise, which is why the metric carries the weight rather than the definition Quotation withheld because this record is not cleared for redistribution. contradicts → statistical parity as a notion of fairness Quotation withheld because this record is not cleared for redistribution. builds-on → differential privacy (tools applied to fairness) Quotation withheld because this record is not cleared for redistribution. extends → differential privacy [Dwo06, DMNS06] Quotation withheld because this record is not cleared for redistribution. builds-on → the exponential mechanism [MT07] Quotation withheld because this record is not cleared for redistribution. builds-on → the Earthmover relaxation for metric labeling and 0-extension [KT02, CKNZ04] Quotation withheld because this record is not cleared for redistribution. statistical parity —same-as→ group fairness statistical parity group fairness strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. Lipschitz condition —causes→ statistical parity Lipschitz condition statistical parity holds when: only when the Earthmover distance between the two groups is small · strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. An algorithm maximizes utility subject to the fairness constraint that similar individuals are treated similarly. Quotation withheld because this record is not cleared for redistribution. An adaptation guarantees statistical parity: the demographics of those receiving any classification match the demographics of the underlying population. Quotation withheld because this record is not cleared for redistribution. task-specific metric task-specific metric Quotation withheld because this record is not cleared for redistribution. fairness constraint fairness constraint Quotation withheld because this record is not cleared for redistribution. utility utility Quotation withheld because this record is not cleared for redistribution. fairness constraint fairness constraint Quotation withheld because this record is not cleared for redistribution. statistical parity statistical parity Quotation withheld because this record is not cleared for redistribution. Lipschitz condition on the classifier Lipschitz condition on the classifier Quotation withheld because this record is not cleared for redistribution. statistical-significance statistical-significance Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1145/2090236.2090255.