All supervised models should, if possible, return attribute weights

yzan

All supervised operators should, if meaningful, return attribute weights representing the feature importance. If nothing else a decision tree and perceptron could get it.

sgenzer

hello @yzan - can you please give us an example to replicate?

 

Scott

yzan

An example of a supervised operator, which returns attribute weights, is "Generalized Linear Model".

 

The calculation of weights for a decision tree:

1. It is possible to simply return a vector, which can take values {0,1} based on whether the attribute is used in the tree, or not.
2. Or http://scikit-learn.org/stable/modules/generated/sklearn.tree.DecisionTreeClassifier.html
3. Or http://support.sas.com/documentation/cdl/en/stathpug/68163/HTML/default/viewer.htm#stathpug_hpsplit_details30.htm
4. Or http://support.sas.com/documentation/onlinedoc/miner/em43/allproc.pdf (pages 54-56).
   

For a perceptron, the returned attribute weights could correspond to the weights of the perceptron (they are already visible in the "model" output, but they are not immediattely passable to operators like "Select by Weights").

sgenzer

thanks for that, @yzan.  Just heard back from dev team that this is coming soon. 

Scott

 

sgenzer

yzan

Possibly even "Deep Learning" could return attribute weights as the backend H2O implementation provides this information and other algorithms from H2O, like GLM and GBT, already output attribute weights.

CraigBostonUSA

Update: As of version 8.0 Decision Tree and Random Forest now provide a new port that outputs feature weights.

 

https://docs.rapidminer.com/la

---

@yzan wrote:

All supervised operators should, if meaningful, return attribute weights representing the feature importance. If nothing else a decision tree and perceptron could get it.

---

test/studio/releases/changes-8.0.0.html?_ga=2.83072976.793993492.1515416834-774805979.1445867999

sgenzer

?