Peer Review #2 of "Model independent feature attributions: Shapley values that uncover non-linear dependencies (v0.1)" [peer_review]

G Erion
2021 unpublished
Shapley values have become increasingly popular in the machine learning literature, thanks to their attractive axiomatisation, flexibility, and uniqueness in satisfying certain notions of 'fairness'. The flexibility arises from the myriad potential forms of the Shapley value game formulation. Amongst the consequences of this flexibility is that there are now many types of Shapley values being discussed, with such variety being a source of potential misunderstanding. To the best of our
more » ... all existing game formulations in the machine learning and statistics literature fall into a category, which we name the model-dependent category of game formulations. In this work, we consider an alternative and novel formulation which leads to the first instance of what we call model-independent Shapley values. These Shapley values use a measure of non-linear dependence as the characteristic function. The strength of these Shapley values is in their ability to uncover and attribute non-linear dependencies amongst features. We introduce and demonstrate the use of the energy distance correlations, affine-invariant distance correlation, and Hilbert-Schmidt independence criterion as Shapley value characteristic functions. In particular, we demonstrate their potential value for exploratory data analysis and model diagnostics. We conclude with an interesting expository application to a medical survey data set. PeerJ Comput. Sci. reviewing PDF | ( ABSTRACT 11 Shapley values have become increasingly popular in the machine learning literature, thanks to their attractive axiomatisation, flexibility, and uniqueness in satisfying certain notions of 'fairness'. The flexibility arises from the myriad potential forms of the Shapley value game formulation. Amongst the consequences of this flexibility is that there are now many types of Shapley values being discussed, with such variety being a source of potential misunderstanding. To the best of our knowledge, all existing game formulations in the machine learning and statistics literature fall into a category, which we name the model-dependent category of game formulations. In this work, we consider an alternative and novel formulation which leads to the first instance of what we call model-independent Shapley values. These Shapley values use a measure of non-linear dependence as the characteristic function. The strength of these Shapley values is in their ability to uncover and attribute non-linear dependencies amongst features. We introduce and demonstrate the use of the energy distance correlations, affine-invariant distance correlation, and Hilbert-Schmidt independence criterion as Shapley value characteristic functions. In particular, we demonstrate their potential value for exploratory data analysis and model diagnostics. We conclude with an interesting expository application to a medical survey data set. 12 13 14 15 16 17 18 19 20 21 22 23 24 25 Shapley values that uncover non-linear dependencies (Sunnies) is, to the best of our knowledge, the 36 only Shapley-based feature importance method that falls into the model-independent category. In this 37 category, feature importance scores attempt to determine what is a priori important, in the sense of 38 understanding the partial dependence structures within the joint distribution describing the DGP. We show 39 that these methods that generate model-independent feature importance scores can appropriately be used 40 as model diagnostic procedures, as well as procedures for exploratory data analysis. 41 Existing methods in the model-dependent category, on the other hand, seek to uncover what is 42 perceived as important by the model (or class of models), either with regards to a performance measure 43 (e.g., a goodness-of-fit measure) or for measuring local influences on model predictions. Model-dependent 44 definitions of feature importance scores can be distinguished further according as to whether they depend 45 PeerJ Comput. Sci. reviewing PDF | ( Manuscript to be reviewed Computer Science on a fitted (i.e., trained) model or on an unfitted class of models. We refer to these as within-model scores 46 and between-model scores, respectively. This distinction is important, since the objectives are markedly 47
doi:10.7287/peerj-cs.582v0.1/reviews/2 fatcat:zqnbks3qyzbinnjkshwxryr4xe