Understanding Variable Effects from Black Box Prediction: Quantifying Effects in Tree Ensembles Using Partial Dependence
Keywords:
Bagging, Bootstrap, Data Mining, Random Forest, Partial DependenceAbstract
Characterizing the impact of one or more factors on a result is often of scientific interest. Although data mining techniques like random forests are adaptable substitutes for traditional parametric models, their lack of interpretability stems from the inability to quantify variable effects in a manner that is substantively relevant. In this study, we offer a partial dependency approach for quantifying variable impacts that averages over the effects of all other variables and yields an estimate that may be understood as the influence on the response for a one unit change in the predictor. Most significantly, the method eliminates issues with model misspecification and implementation difficulties in high-dimensional situations that arise with other methods (e.g., multiple linear regression). We provide a technique for creating a point estimate of this impact magnitude and assess it via simulation. Additionally, we suggest and assess interval estimates using a non-parametric bootstrap. Data used to forecast the age of abalone is used to demonstrate the approach.