3
$\begingroup$

I saw a similar question was asked a few years ago, maybe there are updates on that.

I would like to have a way to explain the decisions generated by random forest, possibly in a single tree. I tried to evaluate each tree and select the best performing one, but using this method I get (as expected) a lot of variation. Is there a way to generalise a random forest which would produce more consistency?

many thanks,

$\endgroup$
  • 1
    $\begingroup$ You generalise a random forest by treating it as... a forest. That is, you take the average prediction from all the trees, rather than from just one. $\endgroup$ – Hong Ooi Sep 24 '14 at 17:14
  • $\begingroup$ It's really hard to understand how the forest makes decisions if you want to explain that to the user. I'm after a method that would make it more human readable. $\endgroup$ – jamborta Sep 25 '14 at 8:11
2
$\begingroup$

How to make otherwise black-box models more interpretable is a great question that I often struggle with. Do any of these related questions help?

You mentioned a lot of variation between trees; I suspect that - by definition - no single tree will be a good representation of the forest. But if the goal is purely to justify/interpret/visualize how the RF makes predictions, there may be simple alternatives. If the number of features is small you could chart how the RF's predictions change over the dataspace. (I've had some success using 2-D and 3-D plots for this with a clinical audience). If the number of features is (slightly) bigger, you could provide a "what-if?" analysis, where users can vary the input values and see the effect on output. And you can always measure the relative contributions of each feature.

Anyway, see some great answers to the above questions for more ideas. I think this is certainly an important question.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.