HpForest, response curve by variable. Possible?

[SBSakura]

Hi all,

I m not sure to post in the right section...i hope!

I work with SAS miner 12.3 and use Random forest node to predict biotope presence in my thesis. I understand and use it easily but i would like to have more information about the model. I have a dataset with one binary variable (presence or absence of the biotope) associated to 50 quantitatives variables.

Is it possible to compute graph of the curve response of presence for each quantitatives variables?

anyone has a idea?

Thank you,

William

[rayIII]

So you want to plot the predicted probabillity of presence/absense against each interval model input? 

Modeling nodes, including the HP Forest node, add predictions as P_ columns to your training table. So one way is to connect  a Graph Explore node to your modeling node, then create scatterplots of p_biotope against each input.

This could be automated too, which the %em_report macro. 

Ray

[SBSakura]

Thank raylll for the reply,

It s not exactly what i want, i want to create graph after running hpforest with the "contribution" of each variable in the prediction (so about 50 graph). example:

1) I have a dataset with binary variable biotope A (presence/absence) and other explicatives variables (like forest_cover)

2) i run hpforest

3)i have the variable importance in the hpforest result (we can say forest_cover is first one)

4) connect hp explore node to create a graph where i can have the response curve of the forest_cover variable. I mean that i just want to have the impact of this variable on the model prediction without considering the other.

Like

Y prob presence (but only for this variable) and X range of forest_cover

where we can see for example that the model choose predicts favorable at the value 5 but unfavorable at value 25 ...

I hope it s not too confused :s

[JasonXin]

Hi,

Your question I think is about Segment profile using EM. You can look into Segment Profile node under ASSESS? group. Also there is a youtube video by my SAS colleague on the subject:

https://www.youtube.com/watch?v=Cm0fDWHpbwg

Hope this helps?