Package/function for neural net-like data transformation


#1

Are there any packages and/or related functions for data transformation a-la a neural network? By this, I mean something like a package called data_transform (which I made up off the top of my head) that, when used like this:

data_transform(mydata)

will give natural logarithms, weird interactions with all of the other variables, etc. for all of the numeric variables in the data.

I know relying on a package/function like that would be incredibly bad for research purposes – I’m just curious if one exists


#2

I don’t know any such package, but there is a product from h2o.ai called Driverless AI that does something like that and more. I don’t think they have R client (yet?), but as an exploratory stage you can use even interactively.


#3

I watched a recording of a demo interfacing with R just this week (company internal so cannot share). I think this may have used Driverless AI, but it may well have been one of the other h2o components.


#4

An auto-encoder does something similar to what you’re asking about, i.e. it generates an encoding of the data that includes interactions and transformation.

For a practical example of this, see the article Predicting Fraud with Autoencoders and Keras.

The article points to several resources, including Chapter 14: Autoencoders of the book “Deep Learning” by Goodfellow, Bengio et al.

More examples at:


#5

It may have been their general h2o thingy with h2o.automl.

To be honest, I don’t see why they have 2 products doing more or less the same thing, but AutoML is available in R, while Driverless AI is not (and not even in a pipeline - pdf, p.5)


#6

Take a look at the recipes package.