Ikibazo gihuza ibibazo · Hagati Urwego
Ikibazo cyo Kuringaniza Uburinganire
Explores individual counterfactual predictions, causal assumptions, protected-attribute changes, and limits of using one fairness measure as proof.
Ikibazo 1 Bya 8
Ni irihe gereranya rishingiye ku butabera budahwitse?
Reba buri kibazo8 ibibazo
Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.
- Ni irihe gereranya rishingiye ku butabera budahwitse?
- Only group average accuracy
- The same individual’s prediction in the observed world and a modeled world with a different protected attribute
- Only the percentage selected in each group
- Only model performance on a test set
- Ni ikihe gikoresho cyo kwerekana icyitegererezo umwimerere wo kurwanya-uburinganire ukoresha?
- A random split of labels
- A cryptographic hash
- A survey with no causal structure
- Structural causal models
- Bigenda bite kuranga ikiranga mukizamini kinyuranyo?
- The analyst specifies an intervention in the structural causal model, changes the protected attribute and evaluates the same person’s prediction
- It is replaced with a group average
- It is always deleted from the dataset without analysis
- It is used only to choose a test threshold
- Ni ukubera iki ubutabera budahuye bushobora kumenya ibibazo impuzandengo y'itsinda rishobora kubura?
- It measures only training speed
- It only tests average error rates
- It examines how a protected-attribute change could alter a particular person’s prediction
- It ignores individual decisions
- Nibihe bitekerezo bigabanya gusobanura ibisubizo bivuguruzanya?
- It requires no assumptions
- It always identifies the true causal graph
- It proves legal compliance
- Results depend on causal-model assumptions and pathway judgments
- Can a system be counterfactually fair under one model and still raise other fairness concerns?
- No, it guarantees every fairness property
- Yes; the criterion does not guarantee equal group outcomes or satisfy every other metric
- No, if it passes, it proves no harm exists
- Yes, but only when there are no protected attributes
- Why should an auditor document which causal pathways are allowed?
- It changes the model’s software license
- The law requires a graph for every model worldwide
- Those choices encode normative assumptions that affect the counterfactual result
- It eliminates data-quality problems
- How should a team interpret a counterfactual-fairness score?
- Evidence conditional on a specified causal model, to be considered with other tests
- A legal decision by a regulator
- A universal certificate of fairness
- A guarantee that a model has no disparate impact