Page MenuHomePhabricator

Include specific user groups in the trwiki edit quality model
Closed, ResolvedPublic

Description

@Superyetkin suggested that some user groups should be whitelisted in the edit quality predictions. List them out, include a feature in the models for them and retrain.

Trwiki specific features live here: https://github.com/wiki-ai/editquality/blob/master/editquality/feature_lists/trwiki.py

We gave a general set of features for user-rights here: https://github.com/wiki-ai/editquality/blob/master/editquality/feature_lists/mediawiki.py#L16

You'll likely want to replicate this pattern inside of the trwiki.py file for the specific "whitelist" user and then retrain the model. To do that, log into ores-compute-01.eqiad.wmflabs and rebuild the trwiki models.

Event Timeline

@Superyetkin, what were the user groups who you think should be excluded?

Halfak triaged this task as Medium priority.Jul 15 2016, 3:08 PM
Halfak moved this task from Unsorted to New development on the Machine-Learning-Team board.
Halfak moved this task from New development to Maintenance/cleanup on the Machine-Learning-Team board.

what were the user groups who you think should be excluded?

"whitelist" group (and those that have the derived privileges) should be excluded.

Is there anyone who can rebuild the model for trwiki?

Yup. We're currently bogged down a bit by some operations work. But we have assigned this task and we should be able to pick it up in the next week. Sorry for the delay and thanks for your patience and collaboration.

@Superyetkin: Hey, I'm going to pick this up and do it. I checked and trwiki doesn't have "whitelist" group neither the right. here's list of them. Can you point out to me what user group do you mean? I can think of autoreview and patroller. Do you mean those?

Yes, the correct group name is "autoreview". The change would also work for the "patroller" group. Thanks.

We still see some edits made by users with "autoreview" rights being flagged as harmful. Here is an example.

It's fixed but it's not deployed yet. We will deploy it possibly this week and also bear in mind we are improving our system to know the rights and possibly don't show the "r" flag for them but we can't eliminate all cases. It's AI after all.