h2o deep features

[mers...@gmail.com]

Hi All,

I am trying to carry out feature selection using deep learning in the h20 program. I tried following the example on http://rpackages.ianhowson.com/cran/h2o/man/h2o.deepfeatures.html and i have problems interpreting the outcomes of deep features selected.

The output ( for example prostate.deepfeatures_layer1) for the run was a dataframe with 100 columns (DF.C1 to DF.C100)and the response type of each sample. An excerpt of the output is shown as follows.

============SAMPLE OUTPUT of Deepfeature==============
CAPSULE DF.C1 DF.C2 DF.C3 ...
1 0 0.6168736 0.1234923 0.4583036 ...
2 0 0.9216827 0.6870487 0.3447747 ...
3 0 1.4295688 0.8358843 0.1626315 ...
4 0 0.0000000 1.3041775 0.0000000 ...
...
...
...
======================================================

I was wondering if i can map these columns (DF.C1 to DF.C100) back to the original dataset and determine which features (AGE,RACE,DPROS,DCAPS,PSA,VO;,GLEASON) in the prostate dataset are important?

Thanks

[arno....@gmail.com]

Hi,

You are probably interested in setting variable_importances=T. You will get feature importances out that are computed by looking at the weights connecting the input features to the first two hidden layers.

See
http://learn.h2o.ai/content/hands-on_training/deep_learning.html
under Variable Importances.

Hope this helps,
Arno

[ivo....@gmail.com]

As I understand h2o.deepfeatures gives the activation of the hidden layer for some input, not the actual features. I would like to get the weight matrix to visualize the features. How can I do that?

[Arno Candel]

Ivo,
You can use h2o-dev and set export_weights_and_biases=T during training, then access the weights from R:

https://github.com/h2oai/h2o-dev/blob/master/h2o-r/tests/testdir_algos/deeplearning/runit_deeplearning_weights_and_biases.R

Hope this helps,
Arno

> On Apr 26, 2015, at 12:51 PM, ivo....@gmail.com wrote:
>
> As I understand h2o.deepfeatures gives the activation of the hidden layer for some input, not the actual features. I would like to get the weight matrix to visualize the features. How can I do that?
>

> --
> You received this message because you are subscribed to a topic in the Google Groups "H2O & Open Source Scalable Machine Learning - h2ostream" group.
> To unsubscribe from this topic, visit https://groups.google.com/d/topic/h2ostream/TB8URytw5y8/unsubscribe.
> To unsubscribe from this group and all its topics, send an email to h2ostream+...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[suneet...@gmail.com]

Hi Arno,

I would like to extract the features by using deep learning and classify it using random forest.

I appreciate if you could able to help me in this.

Thanks in advance
suneetha

[Erin LeDell]

Check out the documentation for h2o.deepfeatures and/or this Deep
Learning tutorial (it's in R):
http://htmlpreview.github.io/?https://github.com/ledell/sldm4-h2o/blob/master/sldm4-deeplearning-h2o.html

-Erin

--
Erin LeDell Ph.D.
Statistician & Machine Learning Scientist | H2O.ai