Page MenuHomePhabricator

[Epic] Explore disparate impacts of damage detection and goodfaith prediction on anons and newcomers.
Closed, ResolvedPublic

Description

We have been receiving reports of false positives in ORES' guesses. It looks like ORES is very skeptical of edits by anonymous editors as well as newcomers. We should explore this problem and see if we can address it.

See:

Event Timeline

Halfak claimed this task.
Halfak raised the priority of this task from to Needs Triage.
Halfak updated the task description. (Show Details)
Halfak moved this task to Backlog on the Machine-Learning-Team (Active Tasks) board.
Halfak added a subscriber: Halfak.
Halfak set Security to None.
Halfak added a subscriber: He7d3r.
ori renamed this task from Explore disperate impacts of damage detection and goodfaith prediction on anons and newcomers. to Explore disparate impacts of damage detection and goodfaith prediction on anons and newcomers. .Dec 2 2015, 10:40 PM

So, I've re-trained all of our edit quality models (except Wikidata) without user.age or user.is_anon. Here's the difference.

wikimodelcurrent AUCno-user AUCdiff
dewikireverted0.9000.792-0.108
enwikireverted0.8350.795-0.040
enwikidamaging0.9010.818-0.083
enwikigoodfaith0.8960.841-0.055
eswikireverted0.8800.849-0.031
fawikireverted0.9130.835-0.078
fawikidamaging0.9510.920-0.031
fawikigoodfaith0.9610.897-0.064
frwikireverted0.9290.846-0.083
hewikireverted0.8740.800-0.074
idwikireverted0.9350.903-0.032
itwikireverted0.9050.850-0.055
nlwikireverted0.9330.831-0.102
ptwikireverted0.8940.812-0.082
ptwikidamaging0.9130.848-0.065
ptwikigoodfaith0.9230.863-0.060
trwikireverted0.8850.809-0.076
trwikidamaging0.8920.798-0.094
trwikigoodfaith0.8990.795-0.104
viwikireverted0.9050.841-0.064

I think that we'll want to compare these models against a set of anon false-positives so that we can assess whether we've addressed the disparity. I think that we should consider running a public discussion about whether or not to switch to the *no-user* models. I'll be happy to layout the tradeoffs and advocate for the switch.

One sad note is that dropping these features means we couldn't claim to be matching the state of the art, however I think that this is good incentive to explore new strategies for improving our signal in other ways.

We talked about this at our most recent revscoring meeting. Here are my notes:

  • This kind of disparate impact is more critical when there's a bot automatically reverting. As we incorporate more human judgement, these issues likely lessen, but do not go away.
  • We could host two models -- one that includes user-features and one that does not. We'd need to change ORES's architecture to support this nicely.
  • @Halfak will summarize the tradeoffs in a post on :m:Research talk:Revscoring
Halfak renamed this task from Explore disparate impacts of damage detection and goodfaith prediction on anons and newcomers. to [Epic] Explore disparate impacts of damage detection and goodfaith prediction on anons and newcomers. .Dec 17 2015, 11:40 PM
Halfak moved this task from Staged to In Progress on the Research board.

I want to bring revscoring up to 1.0.0 (or nearly) so that we can do a good job of making sure we're getting the most signal we can before going back to our users and advocating for a less-fit model.

I ran a test with the new term frequency features and was able to bring the AUC of enwiki damage detection back up to .88 AUC. More testing is needed and other wikis, of course.

It looks like I can drop user.is_anon and user.age from the wikidatawiki models and maintain 0.95 AUC

Marking this as resolved. Presented this at Research Showcase Aug 2016. See T143275