Ikibazo gihuza ibibazo · Biragoye Urwego
Ibipimo bya AI Uburinganire: Uburinganire bwa Demokarasi Kuringaniza Ikibazo
Gerageza uburinganire bwabaturage, amahirwe angana no kunganya ibintu hamwe nurugero rwinguzanyo rwakazi, ibisubizo bidashoboka, nibikoresho bya Fairlearn na AIF360.
Ikibazo 1 Bya 8
Mu ngero z'inguzanyo ziyobora, itsinda A rifite igipimo cyo gutoranya 40% naho itsinda B rifite igipimo cya 20%. Nibihe bipimo bitubahiriza?
Reba buri kibazo8 ibibazo
Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.
- Mu ngero z'inguzanyo ziyobora, itsinda A rifite igipimo cyo gutoranya 40% naho itsinda B rifite igipimo cya 20%. Nibihe bipimo bitubahiriza?
- Equal opportunity
- Demographic parity
- Igenamiterere
- Individual fairness
- Amahirwe angana arasaba ingano ingana gute mumatsinda?
- False positive rate
- Selection rate
- True positive rate
- Positive predictive value
- Niki kuringaniza ibingana byongeraho amahirwe angana?
- Equal false positive rates
- Equal selection rates
- Equal group sizes
- Equal base rates
- Kurugero, itsinda B ryikigereranyo cyibinyoma ni 6 kuri 80. Kuki ari 80?
- It is the number of approved applicants
- It is the total population of both groups
- It is 100 minus the 20 approved applicants
- It is the number of group B applicants who would not repay
- Niki kidashoboka ibisubizo biva kuri Chouldechova na Kleinberg nabandi. vuga?
- No model can be accurate for more than one group
- When base rates differ, an imperfect classifier cannot satisfy predictive parity or calibration together with equal false positive and false negative rates
- Demographic parity always implies equalized odds
- Fairness metrics apply only to linear models
- Why does removing the protected attribute from the inputs not guarantee fairness?
- Models always add the attribute back automatically
- Laws require the attribute to be included
- Correlated proxies such as zip code can carry the same information
- Removing features always lowers accuracy to zero
- Which technique chooses group-specific thresholds, sometimes randomized, to equalize true and false positive rates?
- Hardt et al.'s post-processing, available as Fairlearn's ThresholdOptimizer
- AIF360 Reweighing
- Kureka guta
- Isesengura ryibanze
- Why should fairness metrics be reported with confidence intervals and for intersectional subgroups?
- Because the four-fifths rule requires exactly three decimal places
- Because intersectional groups always have identical rates
- Because Fairlearn cannot calculate point estimates
- Because small groups make rate estimates noisy, and a model can look fair on single attributes while failing a combination of them