Random forest explainability using counterfactual sets

A Survey of Contrastive and Counterfactual Explanation Generation Methods for Explainable Artificial Intelligence

“…Thus, a counterfactual is deemed justified if it can be connected to an associated ground-truth data instance without crossing the decision boundary. Fernández et al introduce the notion of counterfactual sets to enhance counterfactual diversity [81]. They explain random forest predictions by fusing different tree predictors so that the resulting counterfactual set contains the most relevant counterfactual.…”

“…Both types of evaluation methods highlight the features supporting evidence as formulated in the contrastive explanation on the basis of the values that a perturbation variable takes on. Fernández et al evaluate counterfactuals in terms of the average of the pairwise distances based on the feature type and the percentage of valid counterfactuals [81]. Mothilal et al stress that counterfactuals should be evaluated in terms of validity (i.e., whether a generated counterfactual really leads to a different outcome), proximity (i.e., feature-wise distance between the original and counterfactual samples), sparsity (i.e., number of features differing in the original and counterfactual samples), and diversity (i.e., feature-wise distance between each pair of counterfactuals) [105].…”

“…Indeed, contfactuals are particularly suitable for informing the end-user why a given data example is assigned a particular class label. Thus, the outlined classification-oriented frameworks are evaluated on classifiers based on logistic regression [55], [136], [153], [158], decision trees [46], [80], [122], [140], [150], [155], [159], gradient boosted decision trees [147], support vector machines [131], [138], [146], random forests [81], [86], [142]- [144], neural networks [6], [48], [49], [91], [129], [130], [133], [135], [139], [141], [145], [148], [151], or combinations of these [100], [105], [134], [152], [154], [160]. In three studies [67], [128], [137], the classifiers used in the experiments are not specified.…”

“…Poyiadzi et al denote a counterfactual to be ''the new state of the object'' [49]. Nevertheless, the most commonly acceptable definition of a contfactual in the observed RQ 3 -related studies states that a contfactual explanation is a set of minimal feature modifications that makes the model change the prediction [81], [105], [122].…”

“…et al refer to a counterfactual as an effective type of explainable ML technique that explains predictions by describing the changes needed in a sample to flip the outcome of the prediction[81]. More precisely, Fernández et al define a counterfactual for classification tasks as a ''hypothetical instance similar to an example whose explanation is of interest but with different predicted class''[80].…”