Seeing the wood for the trees: Predictive margins for random forests

Recursive partitioning techniques such as classification trees and random forests offer a number of attractive features to corpus data analysts. However, the way in which these models are typically reported – a decision tree and/or set of variable importance scores – usually offers insufficient information to language researchers. Thus, we are usually interested in the (form of) relationship between (multiple) predictors and the outcome, which can be difficult to read from individual trees or variable importance scores. This paper develops predictive margins as an interpretative approach to ensemble techniques such as random forests. Predictive margins are model summaries in the form of adjusted predictions, which allow us to extract from a model the equivalent of what are sometimes referred to as main effects and interaction effects. This provides a clearer picture of patterns in the data and allows us to query a model on potential non-linear associations and interactions among predictor variables. Further, the fact that ensemble methods generate a batch of predictions for a specific condition allows us to obtain indications of statistical uncertainty. The present paper outlines the general strategy for forming predictive margins and addresses methodological issues from an explicitly (corpus) linguistic perspective. We illustrate the technique using data on the English genitive alternation (Grafmiller 2014) and provide an online tutorial on the implementation of predictive margins in R.