What are the plagiarism rules taking in to account the whole point of the exercise is to recreate their research. (E.g. are we allowed to just copy the code and make only tiny modifications?)
In the paper, the 'explanation' is given by a vector δ, based on a counterfactual world in which a class x is defined by t(x) = x + δ, which is simply a translation of all points belonging to the class. Using only translations leads to bad results if regarding two groups with different variances. Thus, we would like to implement t(x) = λx + δ (scaling the cluster points) and t(x) = Mx + δ (transforming the cluster points by a matrix). Do you think this is a good idea? (Also see question 3 and 4)
t(x) = x + δ is an explainable counterfactual because δ is sparse and corresponds to human-understandable features, causing an explanation of the form: "Group A is different from Group B, but would have been the same if Group A would be 8 degrees higher in temperature and 10kg lower in mass" (corresponding in this example to a δ similar to [0, 0.01, 0, 8, 0, ..., -10, 0.02]). But, how can we explain t(x) = λx + δ and t(x) = Mx + δ then?!
Is it fine if our additional research to the paper yields better results, but causes us to not being able to "explain" the results anymore, and just use the additional research we did to emphasize the trade-off between explainability and a good-working algorithm?