Orange Forum • View topic - Index of Tree in the ouptput in Random forest

Index of Tree in the ouptput in Random forest

A place to ask questions about methods in Orange and how they are used and other general support.

Index of Tree in the ouptput in Random forest

Postby collinan » Tue Jul 13, 2010 21:22

Hi Orangers,
I have a basic questions. How do we properly choose the index of tree in the output widget within Random Forest? Is the index "0" corresponds to the best tree?
Many thanks,
Collinan.

Postby marko » Tue Jul 13, 2010 22:28

The Random Forests widget produces classifiers which use all the generated trees. The index parameter was meant to be used for viewing of individual trees.

Trees are added to a list as they are generated and the index only denotes the position in that list. They are not sorted according to quality.

Postby collinan » Wed Jul 14, 2010 0:48

Many Thanks for your explanation Marko.

My question is now: is it possible to represent results of a Random Forest (several trees) by means of a single decison tree (aiming to draw a classification map)?

Another question, is it possible to know the contribution (% of variance explained) of selected splitting variables?

Contribution of splitting variables in decison trees

Postby collinan » Thu Jul 15, 2010 21:55

Actually, my last question seems to be crucial for understanding Random Forest behaviours of splitting variables.
So, how do we know the contribution (% of variance explained) of selected splitting variables?

Postby marko » Fri Jul 16, 2010 6:01

Well, it is quite unlikely that a single tree would produce the same results as random forests. If you want models that are easily interpretable, you should use of the other classification techniques such as naive bayes, classification trees or logistic regression.

You can see the influences of individual attributes for a tree in a built random forest by connecting the "Choosen Tree" output to "Classification Tree Viewer" or "Classification Tree Graph" widgets.

If you would like to see the contributions of individual attributes for the whole forest, you will need to use scripting: check MeasureAttribute_randomForests class in orngEnsemble module.

Postby collinan » Fri Jul 16, 2010 20:24

Thanks a bunch, Marko.
I therefore have to write some codes.
Concerning widgets, what do you mean by "the influences of individual attributes"? Are there some quantitative ouptuts?

Postby marko » Sat Jul 17, 2010 7:48

Yes, you can see the distribution of classes after each split.


Return to Questions & Support



cron