Feature selection algorithm

[FranckG]

Hi all,

I built a huge datamart (more than 8 000 fields) and now I'd like to build a predictive model on a binary target.

I'd like to know if there is an algorithm, like features selection algorithm, that could make a pre-selection of fields. It'd suppress for example all the fields that are not correlated to the target to reduce the number of candidates fields for the modeling.

thanks,

Franck

[NeilR]

Try the Importance Weights tool available in the predictive district here: https://gallery.alteryx.com/#!app/Importance-Weights/56bbd2643df7da08b8fccfe9

[FranckG]

Thanks for your answer.

I downloaded and opened the file 'Importance weights.yxmc' in Alteryx, and do I have to do to use it on my data?

I'm not an expert yet and it's the first time I ue a macro. :)

Franck

[NeilR · Accepted solution]

Once you install the macro (see here), it will appear on your tool palette and you can use it just like any other tool. This particular one requires Java 7 to be installed on your computer.

[FranckG]

Thanks!

[dnoted]

Once I use this tool for a categorical target, how do I got about selecting the right subset of variables? It's giving me 3 different data items as outputs.. 

1. Information gain
2. Gain Ratio
3. Symmetrical Uncertainity

Can I sort Information Gain (1) by descending and take say the top 20? Or do you have any recommendations on determining the threshold?

[jhaugen]

Neil,

I have Java 8 installed on my computer.  Is there anything I have to do differently? Whenever it runs, i get the error ".onLoad failed in loadNamespace() for 'rJava', details: package or namespace load failed for 'FSelector'"

Thanks,

Jake

[NeilR]

Java 8 should be fine - make sure it's x64 though.