I need to select a combination (quadratic and cubic) of features from subset of all namespaces. I want to avoid features which are single. Example:
1 1 '1-1-48112-c|m m_18 |o o_15 |s s_100390 |t t_20 |n n_5283 |d d_10023 |p p_2606 |w w_1 |f f_6.0 |g g_113119
I want to select all features (as per command line of VW I use option -q m: ) which are combination of with namespace 'm'. Importantly, I want to avoid all single feature like t_20, m_18, o_15, etc.
using --ignore params would also remove the combinations, which I don't want.
In the source code it seems ignore namespaces are removed very early before the combinations are made. I am willing to change the source code of VW if needed.
If you want to use the interactions between 2 namespaces, while simultaneously ignoring the 2 namespaces on their own, you are going to need to construct the interaction features on your own and pass them to the model.
It shouldn't be too hard: just contruct a new namespace with the concatenation of features you want, e.g:
1 1 '1-1-48112-c|mynewnamespace t_20_m_18_o_15However, I'm curious why you want to do this? Including the base features plus their quadratic interactions will probably yield similar, if not better, predictive accuracy as taking out the base features.