[M. E. Kaminski, 2019]
E.U.: “right to an explanation” of decisions made
on individuals by algorithms
[FDA Guiding principles]
F.D.A.: “interpretability of the model outputs”
Let be an -person cooperative game with characteristic function
Lloyd S Shapley. A value for n-person games. Contributions to the Theory of Games, 2(28):307–317, 1953.
How important is each player for the outcome of the game?
inputs
responses
Scott Lundberg and Su-In Lee. A Unified Approach to Interpreting Model Predictions, NeurIPS , 2017
Needs of approximations, largely ad-hoc
predictor
We focus on data with certain structure:
Example:
if contains a cross
Fast hierarchical games for image explanations,
Teneggi, Luster & Sulam,
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022
You are telling me that pixels play games?
[Candes et al, 2018]
There exists a procedure that tests for this null and returns a valid p-value,
Large values imply importance in a statistical sense:
Given the Shapley coefficient of any feature
Then
and the p-value obtained for , i.e. ,
From Shapley back to Pearson: Hypothesis Testing via the Shapley Value
J Teneggi, B Bharti, Y Romano, J Sulam
arXiv preprint arXiv:2207.07038
Given the Shapley value for the i-th feature, and
Then, under , is a valid p-value and
prediction should not be correlated with the protected attribute
TPR should be equal for both groups
TPR and FPR should be equal for both groups
(TPRs)
(estimated bias)
Assumption 2:
e.g. if h and f use features that are conditionally independent
* only needs the base rates
Can I make
Under Assumption 2,
where k is an analytical function of the error of h and the base rates
FIFA 20 Data set
A : nationality = {Argentina,England}
Y : salary = {above of median, below of median}
X : player features = {quality score, name}
Estimating and Controlling for Fairness via Sensitive Attribute Predictors
B Bharti, P Yi, J Sulam
arXiv preprint arXiv:2207.12497