[meta] Improve the accuracy of the component model

[suhaibmujahid]

Tasks

Beta Give feedback

1. [model:component] Add features based on file paths in the title and description #4269
   +0
   
2. [model:component] Add sampling techniques to address the imbalanced training dataset #4281
   +0
   
3. [model:component] Experimentation with including Product and Component as a feature #4297
   +0
   
4. Figure out how to handle additional comments in the component model #228
   enhancement +1
5. Consider GeckoView too in Fenix component model #4339
   +0
6. Investigate excluding bugs that received no updates after a component was assigned by bugbot from the training set of the component model #4340
   +0
7. Investigate not training component model on components that haven't been used for a lot of months #2785
   enhancement +1
8. Question: How to mark bugs that are assigned to the wrong component to improve the model? #2227
   +0
9. Evaluate generating multiple component models #224
   +0

[suhaibmujahid]

@sylvestre noticed that the model tends to wrongly classify bugs into the Core::Widget: Gtk component. Here are some examples:

• https://bugzilla.mozilla.org/show_bug.cgi?id=1902245#c2
• https://bugzilla.mozilla.org/show_bug.cgi?id=1853497#c1

[marco-c]

@sylvestre noticed that the model tends to wrongly classify bugs into the Core::Widget: Gtk component. Here are some examples:

* https://bugzilla.mozilla.org/show_bug.cgi?id=1902245#c2

* https://bugzilla.mozilla.org/show_bug.cgi?id=1853497#c1

I bet it's because:

• the "User Agent" mentions Linux;
• Linux is very rare;
• most bugs opened by Linux users are about GTK.

So the "safe bet" for the model is to assign GTK to all bugs mentioning Linux :)

[marco-c]

Another option for improving the model is to use a ML classifier (e.g. using #3761) and then an ensemble of the ML classifier and the xgboost classifier.