Add Stress test : sparse features

[Karim-53]

Normal initial state

linear model
BMI dataset with

Expected: local explanation of a datapoint: BMI feature has a larger variation than gender feature.

Issue

Create an explicand x such that every feature value xi is unique, i.e., it does not occur elsewhere in the training data. (local explanation of a datapoint that is unique in all its xi features)
(A practical implication of Example 3.2 is that the attributions would be very sensitive to noise in the data.)

trigger the issue

If we add a tiny amount of noise, and recompute attributions, then all the features (including BMI and Gender)

The wrong result

Therefore all the variables get equal attributions, even if the function is not symmetric in the variables!

source: https://arxiv.org/pdf/1908.08474.pdf