week 5

Question 1

explainable AI

Answer 1

interpretable (white-box model, can explain the algorithm behind) vs explainable (explain why it works without going into details of the model, which are esoteric anyway)(ad-hoc)
model-based vs model-free(model-agnostic)
local (explain one instance) vs global(all data instances)

Question 2

decision tree

Answer 2

interpretable, model-based, local

Question 3

permutation importance

Answer 3

permute a feature - high loss => important feature, low loss => unimportant feature, show loss discrepancy
global, model-free, explainable(feature summary)
tricky with correlated features, compute on test data

Question 4

counterfactual

Answer 4

explain minimal feature changes that change the prediction
“if x had not happened, then y would have not changed”
local, model-free, explainable(example-based)

Question 5

LIME

Answer 5

fit a linear model for a single data instance by perturbing it, use weights of linear model to explain the most relevant features
local, model-free, explainable(surrogate)