Modelling the trees about randomforest-matlab

[GoogleCodeExporter]

Hi Thomas

sorry for my late reply.

thanks for your helpful comment

will something like the attached m file help? you will need to use graphviz 
(http://www.graphviz.org/) to visualize the tree. After installing graphviz,  
file->new and paste the output. then its F5 to get something like the attached 
figure.

i think the code correctly predicts the tree, though i will debug a bit more 
and see if there is an error or anything.



you will be able to get per tree prediction if you look at 2nd and the 3rd 
output  [Y_new, votes, prediction_per_tree] = classRF_predict(X,model, 
extra_options)

that way its possible to write code to know if permuting a feature increases 
the error rate (aka feature importance)

please do tell me if you need more information.

Original comment by abhirana on 23 Jul 2011 at 10:57

• Changed state: Started

Attachments:

• tutorial_plot_tree.m
• tree_figure.gif

[GoogleCodeExporter]

Hi Abhishek Jaiantilal 

Thank you for a comprehensive answer. It was easy to install and apply when 
logging the command window in MATLAB, however I had trouble plotting the tree 
as vector format, even though the it's set to output as pdf. 
But I got what I needed! 

Thank you once again.
/Thomas

Original comment by [email protected] on 8 Aug 2011 at 12:44

[GoogleCodeExporter]

[deleted comment]

[GoogleCodeExporter]

Hello Sir,

I tried to graphically view one of the tree of RF trained with mixed data (ie. 
both categorical(feature 1-10) having in total 8 categories in each features 
and numerical (features 11-20)) using above given code. 

And the tree obtained is classifying as features 6<=13 or feature 6>13. 
Similarly for feature 8 classifying on the value 12. (the obtained tree is 
attached)

the categorical feature only had values between 1-8 then why it is classifying 
on the value grater than 8. 

Thank You

Original comment by [email protected] on 27 Jun 2013 at 7:53

Attachments:

• .png

[GoogleCodeExporter]

sorry about my reply

if possible could you post the code//data that you used

tree modeling was designed using numerical features. maybe that is the reason.

Original comment by abhirana on 3 Jul 2013 at 11:21