Prescriptive analysis using random forest

Question

Here is a general question in machine learning:

Suppose I have a random forest to predict a failure event in an experiment, the variables consist of both categorical(e.g. type of equipment) and continuous types (time of exposure/amount of reactant).

When it comes to make recommendations to the user (such that he is not going to blow up things), how should I make specific prescriptions, for example use equipment A and add X amount of chemical B?

I was thinking about doing a grid search by building a lattice of potential solutions, and find the optima. But this is expensive and has no guarantee to always return the best solution.

Answer 1

For causal inference you want guarantees that your estimates are unbiased. This paper gives a procedure for generating confidence intervals for random forests, while this paper proves that they are consistent and asymptotically normal. Both papers however deal with regression forests, and not classification forests. It seems like a thoroughly possible extension, but I'm not aware that it has yet been studied in detail.

Answer 2

You could also run straightforward classification and regression trees (CART) analysis using its one-step approach. CART results will inform you which equipment is associated more with accidents, and what level of A is the cutpoint or threshold value above(below) which accidents are mostly seen. I typically use RF to obtain importance scores for features, and obtain a value of class prediction accuracy which is less overfitted.