Evaluate generating multiple component models

[marco-c]

One model for choosing the product, one model to choose the "conflated component", one model to choose the component.

This might help increasing the accuracy (e.g. it's easier to tell that a bug belongs to Core::DOM rather than Core::DOM: Quota Manager).

[yixinsun]

I am thinking of something like 'hierarchical classification'. Did a bit of research and there are 2 ideas:

1. stacking: in my understanding, this is to stack classifiers at different levels paper
2. Hierarchical neural network: model learned from the maximum posterior paper

Are you thinking about something closer to the stacking model?

[marco-c]

Yes.

We have:
Level 1: Product (e.g. Core, Firefox, Toolkit, etc.)
Level 2: Component Group (e.g. Core::DOM*, Core::JavaScript*, etc.)
Level 3: Component

[marco-c]

We could also try https://scikit-learn.org/stable/modules/multiclass.html#multioutput-classification or https://scikit-learn.org/stable/modules/multiclass.html#classifier-chain.

[suhaibmujahid]

See #4172

It appeared to work better when we implemented that for Fenix.