Bias, fairness and explainability quiz

Question 1

What is the key difference between disparate treatment and disparate impact in algorithmic fairness?

Accepted Answer

Disparate treatment is intentional use of a protected attribute; disparate impact is a neutral policy with disproportionate group outcomes

Answer

They are synonyms for the same legal concept

Answer

Disparate impact requires intent while disparate treatment does not

Answer

Disparate treatment applies only to images and disparate impact only to tabular data

Question 2

State the demographic parity (statistical parity) criterion and one major limitation.

Accepted Answer

Equal positive prediction rates across groups, but it ignores the ground-truth label and base rates

Answer

Equal accuracy across groups, ignoring false positives

Answer

Equal calibration across groups, ignoring acceptance rates

Answer

Equal true positive rates across groups, ignoring base rates

Question 3

Define equalized odds and explain how it differs from equal opportunity.

Accepted Answer

Equalized odds equalizes both TPR and FPR; equal opportunity equalizes only TPR

Answer

Equalized odds equalizes acceptance rates; equal opportunity equalizes calibration

Answer

They are identical except equal opportunity also constrains FPR

Answer

Equal opportunity equalizes both TPR and FPR; equalized odds only TPR

Question 4

What does the fairness 'impossibility theorem' (Kleinberg–Chouldechova) state?

Accepted Answer

When base rates differ, calibration and balanced FPR/FNR cannot all hold simultaneously

Answer

All fairness metrics are equivalent and can always be satisfied together

Answer

Calibration and demographic parity are always jointly achievable

Answer

Fairness is impossible unless the protected attribute is removed

Question 5

What is the 'four-fifths' (80%) rule and what fairness metric does it operationalize?

Accepted Answer

A disparate-impact screen flagging selection-rate ratios below 0.8 across groups

Answer

A rule requiring 80% model accuracy on each group

Answer

A calibration threshold of 0.8 across groups

Answer

A requirement that 80% of training data be from minority groups

Question 6

Explain why 'fairness through unawareness' (dropping the protected attribute) is generally insufficient.

Accepted Answer

Correlated proxy features let the model reconstruct the protected attribute despite its removal

Answer

It changes the loss function in an unstable way

Answer

It always reduces accuracy to zero

Answer

It violates demographic parity by definition

Question 7

What is counterfactual fairness?

Accepted Answer

The prediction is unchanged in a counterfactual where the individual's protected attribute is altered, per a causal model

Answer

The prediction equals the average over all groups

Answer

The model is retrained on counterfactual (flipped-label) data

Answer

Predictions are equal for the nearest counterfactual example regardless of causality

Question 8

Contrast pre-processing, in-processing, and post-processing bias-mitigation strategies.

Accepted Answer

Pre = modify data, In = modify training objective, Post = adjust trained outputs

Answer

Pre = adjust outputs, In = modify data, Post = modify objective

Answer

All three only change the threshold

Answer

They differ only in which fairness metric they target

Question 9

How does adversarial debiasing enforce fairness during training?

Accepted Answer

An adversary predicts the protected attribute from outputs; the model is trained to defeat it

Answer

The model adds Gaussian noise to the protected attribute

Answer

Two models vote and the fairer one wins

Answer

It augments data with adversarial perturbations to inputs only

Question 10

What does the reweighing pre-processing technique do?

Accepted Answer

Assigns sample weights making group and label independent under the weighted distribution

Answer

Deletes minority samples to balance classes

Answer

Re-scales features to zero mean per group

Answer

Re-labels misclassified minority instances

Question 11

Are Shapley values for feature attribution guaranteed to be unique, and why?

Accepted Answer

Yes, it is the unique value satisfying efficiency, symmetry, dummy, and additivity

Answer

No, Shapley values are always non-unique

Answer

Yes, but only for linear models

Answer

No, uniqueness holds only when features are independent

Question 12

Why can SHAP/Shapley attributions be misleading when input features are highly correlated?

Accepted Answer

Coalition imputation goes off-manifold or splits credit among correlated features, changing attributions

Answer

Correlated features make SHAP exactly zero

Answer

SHAP cannot run on more than 10 features

Answer

Correlation only affects LIME, not SHAP

Question 13

How does LIME generate a local explanation, and what is its core assumption?

Accepted Answer

Fits a locally weighted interpretable surrogate on perturbed samples, assuming local linearity

Answer

Computes exact gradients of the model globally

Answer

Removes features one at a time and retrains the model

Answer

Builds a global decision tree mimicking the model

Question 14

What is the difference between local and global model explanations?

Accepted Answer

Local explains one prediction; global describes overall model behavior across data

Answer

Local is for deep nets; global is for trees

Answer

Local uses gradients; global never uses gradients

Answer

They are identical at scale

Question 15

Why are Partial Dependence Plots (PDPs) unreliable under feature correlation, and what fixes this?

Accepted Answer

PDPs extrapolate to unrealistic feature combinations; ALE plots use local conditional effects to avoid this

Answer

PDPs need labels; ALE does not, so ALE is always wrong

Answer

PDPs only work for linear models; ALE for trees

Answer

Correlation does not affect PDPs at all

Question 16

What property makes Integrated Gradients a principled attribution method for deep networks?

Accepted Answer

It integrates gradients from a baseline to the input and satisfies completeness and sensitivity axioms

Answer

It uses only the final-layer gradient at the input

Answer

It averages SHAP and LIME outputs

Answer

It requires retraining with a fairness penalty

Question 17

What does Grad-CAM produce and what are its inputs?

Accepted Answer

A class-discriminative heatmap from gradient-weighted convolutional feature maps

Answer

A per-pixel SHAP value table

Answer

A surrogate decision tree for the CNN

Answer

A saliency map computed only from input pixels, ignoring conv layers

Question 18

Is attention weight a reliable explanation of a transformer's prediction?

Accepted Answer

No; attention is often not faithful and alternative weights can give the same output

Answer

Yes; attention weights are provably causal explanations

Answer

Yes; attention equals exact Shapley values

Answer

No; attention is never even plausible to humans

Question 19

Distinguish faithfulness from plausibility in explanation evaluation.

Accepted Answer

Faithfulness = reflects true model reasoning; plausibility = how convincing to humans

Answer

They are the same property under different names

Answer

Faithfulness = human agreement; plausibility = output fidelity

Answer

Plausibility is measurable but faithfulness is not

Question 20

What is a counterfactual explanation and what makes a good one?

Accepted Answer

The minimal, actionable feature change that flips the decision

Answer

A SHAP value for the protected attribute

Answer

A global surrogate model of the classifier

Answer

The training example most similar to the input

Question 21

What are Anchors explanations?

Accepted Answer

High-precision IF-THEN rules giving sufficient conditions that fix the prediction

Answer

Gradient-based saliency maps for text

Answer

A global tree distilled from the model

Answer

Counterfactual edits to flip the prediction

Question 22

What is the well-known pitfall of impurity-based (Gini) feature importance in tree ensembles?

Accepted Answer

It is biased toward high-cardinality/continuous features and computed on training data

Answer

It only works for linear models

Answer

It always equals SHAP values

Answer

It cannot be computed for random forests

Question 23

What is the main weakness of permutation feature importance, and how can it be mitigated?

Accepted Answer

It breaks feature correlations, creating off-manifold inputs; fix by permuting correlated groups or conditional permutation

Answer

It cannot handle classification, only regression

Answer

It requires gradients and fails on trees

Answer

It is only valid on the training set

Question 24

Differentiate sampling bias, measurement (label) bias, and aggregation bias.

Accepted Answer

Sampling = unrepresentative data; measurement = flawed feature/label proxies; aggregation = one model forced across distinct subgroups

Answer

All three refer to label noise

Answer

Sampling = label noise; measurement = class imbalance; aggregation = overfitting

Answer

They are stages of gradient descent

Question 25

How can Simpson's paradox manifest as aggregation bias in fairness analysis?

Accepted Answer

Aggregate trends can reverse within subgroups due to a confounder, so audits must disaggregate

Answer

It means subgroup metrics always equal aggregate metrics

Answer

It only occurs in image data

Answer

It is eliminated by removing the protected attribute

Question 26

What is a feedback loop in a deployed ML system and why is it a fairness concern?

Accepted Answer

Model decisions shape future training data, amplifying initial biases over time

Answer

The loss oscillates and never converges

Answer

Users repeatedly request the same prediction

Answer

Cross-validation folds leak into each other

Question 27

What is predictive parity (calibration within groups) and when does enforcing it conflict with balanced error rates?

Accepted Answer

Equal PPV/calibration across groups; conflicts with balanced FPR/FNR when base rates differ

Answer

Equal recall across groups; never conflicts with anything

Answer

Equal acceptance rates; identical to demographic parity

Answer

Equal accuracy across groups; only relevant for regression

Question 28

What do model cards and datasheets for datasets document, and why do they matter for responsible AI?

Accepted Answer

Standardized transparency docs covering intended use, disaggregated performance, data provenance, and limitations

Answer

Hyperparameter logs only

Answer

Legal contracts transferring model IP

Answer

Automated bias-correction scripts

Question 29

Contrast intrinsic interpretability with post-hoc explainability, including a key trade-off.

Accepted Answer

Intrinsic = transparent by design (faithful); post-hoc = approximate explanations for black boxes (may be unfaithful)

Answer

Intrinsic is always less accurate and post-hoc is always faithful

Answer

They are interchangeable terms

Answer

Post-hoc applies only to linear models

Question 30

What is TCAV (Testing with Concept Activation Vectors) and what does it measure?

Accepted Answer

Quantifies a prediction's sensitivity to a human concept via directional derivatives along a concept activation vector

Answer

Computes per-pixel saliency maps

Answer

Distills the network into a decision tree

Answer

Measures attention entropy across heads

Test yourself