SeterusnyaPanduan seterusnya
Keadilan Berlawanan
Masyarakat
PANDUAN Masyarakat
Several impossibility results show that common statistical fairness criteria can conflict.
Dengar panduan iniAI-diceritakan · ~3 min
For risk scores with different group base rates and imperfect prediction, calibration and two score-balance conditions—mean scores within positive and negative outcome groups—cannot generally all hold exactly; related classifier results concern thresholded error rates.
“The impossibility theorem” is shorthand for several results about incompatible statistical fairness criteria. Kleinberg, Mullainathan and Raghavan’s 2016 paper formalizes three score-level conditions: calibration within groups; balance for the positive class, meaning the average score among people with a positive outcome is equal across groups; and balance for the negative class, meaning the average score among people with a negative outcome is equal across groups. Except in constrained cases—including equal base rates or perfect prediction—one cannot satisfy all three simultaneously. These positive- and negative-class balance conditions compare average scores conditional on outcomes; they are not directly false-positive or false-negative rate parity after thresholding. Calibration means that among people assigned a given score, the observed outcome frequency matches that score within each group. A classifier applies a threshold to a score and assigns a discrete decision. Chouldechova’s related 2017 result addresses a thresholded classifier: with differing outcome prevalence and an imperfect predictor, predictive parity and equal false-positive/false-negative rates cannot both generally hold. That classifier-level error-rate balance is related to, but distinct from, the score-level positive- and negative-class balance conditions in Kleinberg et al. These results depend on the target, base rates, score quality and chosen criteria; they do not establish a universal impossibility of fairness. The theorem does not tell a decision maker which criterion to prioritize, establish that observed labels are valid ground truth, or guarantee that optimizing any one metric produces just outcomes. Context matters: false positives and false negatives may have different costs, scores may be used as rankings rather than probabilities, and group labels or outcomes can themselves reflect structural disadvantage. Woodworth and colleagues study conditions under which one can learn fair representations while navigating trade-offs, further illustrating that conclusions depend on the formal setup. Responsible use requires stating the assumptions and policy choice, reporting resulting harms, and considering procedural and causal approaches beyond the statistical parity metrics.
A risk-score team states whether it prioritizes calibration, equalized error rates or another criterion before comparing groups.
A lending audit reports base rates and false-positive/false-negative rates so stakeholders can see which fairness properties conflict.
A policy maker explains why one metric is prioritized given the decision’s harms instead of claiming a mathematical theorem selected the policy.
A model reviewer tests whether a claimed trade-off applies because group base rates differ and predictions are imperfect.
Fairness research continues to refine conditions for scores, classifiers, multiple groups and uncertain labels. Treat each theorem as scoped to its assumptions, and revisit metric choices when base rates, outcome definitions or decision thresholds change. Keep a dated record of the primary source or study behind each claim and revisit conclusions when new evidence or implementation details emerge. Newer work examines alternative score types, more than two groups and imperfect labels. These extensions do not remove the need to state assumptions and the actual policy objective.
Newsletter percuma
Tiga cerita AI yang disahkan setiap pagi hari bekerja, ditulis dalam Bahasa Inggeris biasa. Percuma selamanya, tiada iklan.
Satu e-mel setiap hari bekerja. Berhenti melanggan dalam satu klik. Kami tidak pernah menjual atau berkongsi alamat anda.
Uji diri anda
Maklum balas segera pada setiap jawapan, dan sijil yang boleh dikongsi dengan ID yang boleh disahkan sebaik sahaja anda lulus kursus.
Sokong pendidikan AI percuma. AI Understanding ialah pertubuhan bukan keuntungan 501(c)(3) — tiada iklan, tiada bayaran dinding, selamanya. Buat derma
Several impossibility results show that common statistical fairness criteria can conflict. For risk scores with different group base rates and imperfect prediction, calibration and two score-balance conditions—mean scores within positive and negative outcome groups—cannot generally all hold exactly; related classifier results concern thresholded error rates.
Kertas ini memformalkan keadaan penentukuran dan keseimbangan untuk kelas positif dan negatif.
Teorem mengenal pasti kes terhalang termasuk kadar asas yang sama atau ramalan yang sempurna.
Penentukuran bermakna skor sepadan dengan frekuensi hasil yang diperhatikan dalam setiap kumpulan.
Kriteria tahap pengelas berkaitan ini membandingkan kadar positif palsu dan negatif palsu merentas kumpulan selepas ambang keputusan; ia berbeza daripada keadaan keseimbangan peringkat skor KMR.
Ketidakserasian timbul di bawah kelaziman kumpulan yang berbeza apabila ramalan tidak sempurna.
Teruskan belajar
Lebih banyak panduan dipilih untuk topik ini
SeterusnyaPanduan seterusnya
Keadilan Berlawanan
Masyarakat