I have an h2o random forest algorithm. I trained the algorithm, tested it, and interpreted the performance. The algorithm is a binary classifier, so it's spitting out 1s and 0s for each record in the test dataset. I reviewed the variable importance plot to determine which fields are contributing the most to the predictions. However, I would like to know which values within those fields are the most important for determining why a record is classified as a 1. How do I do this?
Some ideas. First, you could print out the trees (or export a POJO and go through the java code). This will tell you some of the decision thresholds being used.
Alternatively, partial dependence plots might be what you are after?
The other road you could take is to treat the model as a black box (and once the number of trees in your forest increases, it effectively is), and try something like LIME to explain the model. (Search "h2o" and "LIME" to get some suggestions.)