5.9 When to specify \(\phi\) instead of learning it!

A neural net is the wrong choice on a small tabular regression problem with a few smooth effects. If you already know that price is linear in size above \(1100\) ft\(^2\) and linear in an interaction \(\mathrm{beds}\times\mathrm{wealth}\), type those columns and run OLS or ridge. You will get a standard error, an identifiable coefficient, and a model you can interpret. The net will try to rediscover the same \(\phi_\beta\), but it will not be clear how.

Nets are useful when \(\phi\) is challenging: many inputs, unknown interactions, or a representation you intend to reuse. The next chapter adds a third option. A regression tree will cut if the data want that cut, without requiring a gradient through a corner. Trees and nets are two ways to let the data propose \(\phi\), while in the previous chapter we (as statisticians) proposed \(\phi\).