Do Neural Networks need "compound" features?

Question

Apologies if I haven't got the terminology quite right. I have a question about Neural Networks, and I'm not sure exactly the best way to ask it!

Hypothetically, let's say I have a dataset of houses on the property market. One of the features could be number of bathrooms, and another is floor plan area.

If I were using linear regression, I might also include a compound feature e.g. number of bathrooms / floor plan area. I could come up with all sorts of combinations of feature products, and create higher order polynomials.

My question is - is this necessary for neural networks? Or standard practise? Or useful?

Neural networks are such a black box I'm not sure if it would just "work this out" or not.

If there isn't a simple answer, I would be grateful at least for anything you can tell me about how this phenomenon is referred to. Are they compound features? Or computed features? I really don't know...

Answer 1

composite features can be useful but you probably wouldn't want to use the elements of a composite feature and the composite feature itself as inputs to the same model unless each was meaningful a priori (and even then, maybe not). The net will be able to learn about the summation of two features (and other relationships) without combining them.

a great use for composite features can be when you want to do large scale machine learning and cut down on processing requirements

so not necessary but occasionally useful

one caveat - the nature of a feature might call for a transformation (square root, polynomial transformation) to increase the fit of a simple model but that is a separate issue from combining features into a composite feature