How to determine the range of the feature values used to predict a class using Random Forest Classifier

Question

I am trying to find out what the range of the values of the features are which the random forest classifier uses to predict a certain class.

For example we have the IRIS dataset;

I use the random forest classifier to predict which class of flower species a flower belongs to based on its features, there are 4 features (sepal length, sepal width, petal length, petal width).

I can find out the feature importance and I can visualize the steps the classifier takes by using Graphviz. Now I would like to find out what the range is of for example, sepal length which will result in the prediction being a Setosa, i.e petal length between 0.2 and 0.4 is an indicator that the species is Setosa. I can view this data visually using Graphviz, but I would like a way to store it and analyse it for my entire dataset using 200 estimators. Is there a way to collect and store the data textually, in the image below of the decision tree; if petal-length <= 2.6 then the class is Setosa.

https://images.app.goo.gl/pPK1KsXAMY3z27JW8

I would like to have a data frame similar to this:

        node |   feature     | Samples | Value     | Class 
   --------------------------------------------------------------
   1.    1   |  sepal-length |    23   |  <= 0.2   | Setosa
   2.    3   |  petal-width  |    45   |  <= 0.3   | Versicolor
   3.   ...          ...         ...        ...        ...
   n.    178 |  sepal-width  |    3    |  <= 0.4   | Setosa

Once I have a data frame I would then be able to analyse and see that for example; A setosa flower will have petal-length between 0.1 - 0.3, sepal-length between 0.4-0.7 etc

Would this even be possible? If so, any ideas would be greatly appreciated.

EDIT: I have looked at the decision path for each tree, and while helpful, they do not contain the predicted class, therefore does not help me with what I am trying to do.

I think my only option is to just parse the dot file which I get from the Graphviz function and manually store the information into a data frame.

Answer 1

The RandomForestClassifier of scikit-learn has an attribute called estimators_ which after training is a list of DecisionTreeClassifier instances that together make up the forest.

Now that we can access the individual trees, we take a closer look at the DecisionTreeClassifier instances. Every instance has a tree_ attribute, which contains the actual decision tree and all of the properties you are interested in.

The awesome people of scikit-learn even wrote documentation on how to access the tree's properties.

I understand you are in need of the majority class per node, as in the graphviz visualization, which isn't a standard attribute on the nodes. You mentioned that you might parse the output of graphviz, but maybe you can take a look at the graphviz code!

As you can see at this line in the node_to_str function they take the argmax over the value variable which is defined here. I think that if you combine that with the documentation linked above, you should be able to get the class per node!