Closed tabularaza27 closed 10 months ago
For Ground_truth evaluations, one could use benchmark functions used in optimization like Ackley (or any other known mathematical functions) and fit several linear regression models for several particular ranges of inputs. Weights of linear regression models could be used as ground truth for faithfulness metrics.
Are you planning to add support for regression tasks ? If so, I'd suggest replacing the exact equality condition of the original and perturbed input predictions in the stability metrics (
ŷ_x == ŷ_x'
) to something like|ŷ_x- ŷ_x'|< eps
, where eps is a user-defined range.