Structure Learning

[paroma]

Structure learning using robust PCA via CVXPY.

Overall Implementation:

• there is parent class DependencyLearner whose children differ by the learn_dependencies method
• there is only one subclass currently, which is RPCADependencyLearner. Eventually, we will have others, maybe put a few in contrib/
• user can pass in a DependencyLearner object to LabelModel.train() instead of the junction_tree in the Sigma_O case, and with L_train in the normal case

Issues/Concerns:

• Structure learning works with Sigma_O of size m x m only. What's the best way to ensure this? Current implementation turns higher_order_cliques = False when creating the junction tree
• There's some weirdness with how to pass in either a JunctionTree or DependencyLearner (it works but its not the best implementation/check I think)
• Where do the hyperparameters for lm.train() go? thresh is a common hyperparameter for all dependency learners, but wasn't sure where to pass it in

[ajratner]

Awesome! Will look over soon! First quick question- only working for m x m sigma_O corresponds with binary, ind LFs- so basically this means that SL works if you "start from scratch", but not if you already know some deps? Correct?

[paroma]

i didnt get a notification :( i think if some deps are known, we can "add on" to those. the option to pass in deps still exists!

the second thing is i need to index into whatever sigma_O to create the m x m matrix...if that's easily doable then we can get rid of HIGHER _ ORDER _ CLIQUES = False since it has to be true for prediction anyways

[bhancock8]

What's the next step here? No one is waiting for review still, right?

[ajratner]

This is probably blocked on me finishing a refactor of the LabelModel on this branch- coming soon!!

[paroma]

WIP: working on the 3 view class balance model with current LabelModel implementation in master