Abstract:
|
The goal of this study is to identify important predictors for black box machine learning methods, where the prediction function is highly non-linear, non-additive and cannot be represented by statistical parameters. Thus, such black-box models lack interpretability and it is very difficult to identify ``important'' or significant inputs for an outcome from such models. The main target is to investigate applicability of permutation based approach, proposed by Breiman (2001) and then generalized by Fisher et.al (2019) to the common non-linear and non-additive machine learning techniques. Another aim is to decompose the proposed variable importance metric (VIM) to obtain a causal parameter which is a function of the expected conditional average treatment effect squared over the distribution of treatments for multinomial and continuous treatments. A simulation study was then conducted to check the performance of the estimated VIM using split-sampling techniques using multiple known machine learning methods. The estimation technique of VIM was also evaluated under model mis-specifications.
|