Model interpretability (SHAP, LIME) quiz

Question 1

What theoretical foundation do SHAP values borrow from, and what does this guarantee?

Accepted Answer

Shapley values from cooperative game theory, guaranteeing local accuracy, missingness, and consistency

Answer

Bayesian inference, guaranteeing calibrated posteriors

Answer

Information theory, guaranteeing maximal entropy attributions

Answer

Gradient descent, guaranteeing convergence to global optima

Question 2

In the SHAP framework, what does the 'local accuracy' (efficiency) property require?

Accepted Answer

The sum of SHAP values plus the base value equals the model's prediction

Answer

Each SHAP value is bounded between -1 and 1

Answer

The largest SHAP value equals the prediction

Answer

SHAP values sum to zero across features

Question 3

How does LIME generate an explanation for a single prediction?

Accepted Answer

It fits a locally weighted interpretable surrogate model on perturbed samples around the instance

Answer

It computes exact Shapley values via coalition enumeration

Answer

It backpropagates gradients to input features

Answer

It retrains the model on a subset of features

Question 4

Why is exact computation of Shapley values intractable, and how does KernelSHAP address it?

Accepted Answer

Coalitions grow as 2^M; KernelSHAP uses weighted linear regression on sampled coalitions

Answer

Shapley values require matrix inversion; KernelSHAP uses SVD

Answer

Shapley values need labeled data; KernelSHAP uses self-supervision

Answer

Computation needs the Hessian; KernelSHAP approximates it with finite differences

Question 5

What is the key advantage of TreeSHAP over KernelSHAP for tree-based models?

Accepted Answer

It computes exact SHAP values in polynomial time by exploiting tree structure

Answer

It approximates SHAP values faster but with higher variance

Answer

It only works for shallow trees of depth ≤ 3

Answer

It requires GPU acceleration to be tractable

Question 6

What does a SHAP dependence plot reveal that a global feature-importance bar chart cannot?

Accepted Answer

The nonlinear and interaction structure of a feature's effect across its value range

Answer

The statistical significance of each feature

Answer

The training loss contribution per feature

Answer

The correlation matrix among all features

Question 7

What is the difference between interventional and observational (conditional) SHAP, and why does it matter for correlated features?

Accepted Answer

Interventional uses marginal sampling (true to model); conditional uses conditional sampling (true to data)

Answer

Interventional is for classification; conditional is for regression

Answer

Interventional ignores the base value; conditional includes it

Answer

They are identical for tree models only

Question 8

A stakeholder asks why LIME explanations for the same instance differ between two runs. What is the most likely cause?

Accepted Answer

Stochastic perturbation sampling produces different local datasets each run

Answer

The model weights changed between runs

Answer

LIME caches stale explanations

Answer

SHAP and LIME were mixed in the same call

Question 9

In KernelSHAP and LIME, what role does the kernel/proximity weighting play, and how do they differ?

Accepted Answer

LIME uses a heuristic proximity kernel; KernelSHAP uses the Shapley kernel that yields Shapley values

Answer

Both use identical kernels and differ only in code

Answer

LIME weights by label; KernelSHAP weights by gradient

Answer

Neither uses any weighting

Question 10

What is the SHAP base value (expected value), and how is it computed?

Accepted Answer

The model's expected output over the background dataset, E[f(x)]

Answer

The maximum prediction in the dataset

Answer

The intercept of the original model

Answer

Always zero by definition

Question 11

Why does the choice of background dataset critically affect SHAP explanations?

Accepted Answer

It defines the reference distribution for absent features and the base value

Answer

It only affects plot colors

Answer

It determines the model's learning rate

Answer

It is irrelevant for tree models

Question 12

How do SHAP interaction values extend standard SHAP values?

Accepted Answer

They decompose attributions into main effects plus pairwise interaction effects

Answer

They measure correlation between raw features

Answer

They are SHAP values squared

Answer

They aggregate SHAP values across the dataset

Question 13

A practitioner uses LIME on tabular data and gets nonsensical explanations. What discretization issue could be responsible?

Accepted Answer

Default discretization/binning creates unrealistic perturbations

Answer

LIME cannot handle continuous features at all

Answer

The model was not normalized

Answer

LIME requires one-hot encoding of all features

Question 14

What is the consistency (monotonicity) property of SHAP, and why is it important?

Accepted Answer

If a feature's contribution grows, its SHAP value cannot decrease; ensures reliable comparison

Answer

SHAP values must increase monotonically with feature value

Answer

Predictions must be monotone in every feature

Answer

Larger datasets always give larger SHAP values

Question 15

For deep neural networks on images, which SHAP variants are commonly used and why?

Accepted Answer

DeepSHAP and GradientSHAP, because they use backprop for efficiency on high-dimensional inputs

Answer

KernelSHAP only, because it is exact for images

Answer

TreeSHAP, because CNNs are tree ensembles

Answer

LIME only, because gradients are unavailable

Question 16

What does a SHAP force plot communicate for a single prediction?

Accepted Answer

How each feature pushes the prediction from the base value to the final output

Answer

The model's training curve

Answer

Feature correlations as a heatmap

Answer

The decision boundary in 2D

Question 17

Why can global feature importance derived from mean(|SHAP|) be misleading if interpreted causally?

Accepted Answer

It reflects model behavior, not real-world causal effects

Answer

It is always biased toward continuous features

Answer

It cannot be computed for classifiers

Answer

It double-counts the base value

Question 18

Write Python to compute and summarize SHAP values for an XGBoost model using TreeSHAP.

Accepted Answer

shap.TreeExplainer(model) then explainer(X) and shap.plots.beeswarm

Answer

shap.LinearExplainer for XGBoost

Answer

model.shap_values() built into XGBoost

Answer

shap.KernelExplainer is the only option for trees

Question 19

What is the fundamental difference in scope between LIME and SHAP regarding local vs global explanations?

Accepted Answer

LIME is purely local; SHAP values aggregate consistently into global explanations

Answer

SHAP is purely local; LIME is purely global

Answer

Both are only global

Answer

Both are only local with no aggregation

Question 20

What is the 'missingness' axiom in additive feature attribution, and what does it imply?

Accepted Answer

A feature absent from the instance gets zero attribution

Answer

Missing features get the base value

Answer

Missing features get the maximum SHAP value

Answer

Missingness only applies to categorical features

Question 21

How should SHAP values for a multi-class classifier be interpreted?

Accepted Answer

One attribution set per class, explaining each class's output separately

Answer

A single averaged attribution across all classes

Answer

Only attributions for the predicted class are defined

Answer

SHAP does not support multi-class problems

Question 22

What is a key fidelity-vs-interpretability tradeoff in LIME's surrogate model choice?

Accepted Answer

A linear surrogate is interpretable but loses fidelity where the local surface is nonlinear

Answer

Richer surrogates are always both more faithful and more interpretable

Answer

Fidelity is guaranteed regardless of kernel width

Answer

LIME surrogates are always perfectly faithful

Question 23

For NLP text classifiers, how do LIME and SHAP typically represent features for explanation?

Accepted Answer

As presence/absence of tokens, perturbed by masking words

Answer

As raw embedding gradients only

Answer

As character n-gram frequencies exclusively

Answer

As TF-IDF scores that cannot be perturbed

Question 24

Why might SHAP and LIME disagree on feature importance for the same prediction?

Accepted Answer

They optimize different objectives with different sampling, weighting, and locality

Answer

One uses the training labels and the other does not

Answer

SHAP is random while LIME is deterministic

Answer

They always agree; disagreement indicates a bug

Question 25

What problem does Partition SHAP (PartitionExplainer) solve, and how?

Accepted Answer

It computes Owen values over a feature hierarchy, respecting structure and improving speed

Answer

It partitions the dataset into train/test for validation

Answer

It only works on perfectly independent features

Answer

It replaces the model with a decision tree

Question 26

In a SHAP beeswarm (summary) plot, what do the position, color, and clustering of points convey?

Accepted Answer

Position=impact sign/magnitude, color=feature value, density=distribution of effects

Answer

Position=training epoch, color=class, density=loss

Answer

Position=correlation, color=p-value, density=sample size

Answer

All three encode the same single importance number

Question 27

What is the computational complexity of KernelSHAP with respect to model evaluations, and why is it a concern at scale?

Accepted Answer

Cost scales with coalitions × background size per instance, making it slow at scale

Answer

It is O(1) per instance regardless of features

Answer

It needs only one model evaluation per instance

Answer

Cost is independent of the background dataset

Question 28

How can SHAP be used to detect data leakage or spurious correlations in a model?

Accepted Answer

High implausible attribution to an ID/proxy feature reveals leakage

Answer

SHAP automatically deletes leaky features

Answer

SHAP only detects leakage in linear models

Answer

Leakage cannot be diagnosed with attributions

Question 29

What is a documented vulnerability of LIME and KernelSHAP to adversarial manipulation?

Accepted Answer

Adversaries detect OOD perturbation queries and return innocuous outputs to hide bias

Answer

They leak training data through gradients

Answer

They overflow memory on adversarial inputs

Answer

They cannot be attacked because they are model-agnostic

Question 30

When explaining a linear model, what relationship do SHAP values have to the model coefficients?

Accepted Answer

φi = βi·(xi − E[xi]) for independent features

Answer

φi equals βi exactly regardless of feature value

Answer

φi = βi·xi with no centering

Answer

SHAP is undefined for linear models

Test yourself