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:
| Status | Subtype | Assigned | Task | ||
|---|---|---|---|---|---|
| Resolved | Halfak | T120138 [Epic] Explore disparate impacts of damage detection and goodfaith prediction on anons and newcomers. | |||
| Resolved | Halfak | T122269 [epic] revscoring 1.0.0 | |||
| Resolved | Halfak | T121358 Update french language utilities with expanded badword/informals | |||
| Resolved | Halfak | T121005 Meta datasource/feature refactoring for revscoring | |||
| Resolved | Halfak | T121003 Implement word frequency diff features | |||
| Resolved | None | T118982 hewiki "reverted" model weights strongly against anons |
So, I've re-trained all of our edit quality models (except Wikidata) without user.age or user.is_anon. Here's the difference.
| wiki | model | current AUC | no-user AUC | diff |
|---|---|---|---|---|
| dewiki | reverted | 0.900 | 0.792 | -0.108 |
| enwiki | reverted | 0.835 | 0.795 | -0.040 |
| enwiki | damaging | 0.901 | 0.818 | -0.083 |
| enwiki | goodfaith | 0.896 | 0.841 | -0.055 |
| eswiki | reverted | 0.880 | 0.849 | -0.031 |
| fawiki | reverted | 0.913 | 0.835 | -0.078 |
| fawiki | damaging | 0.951 | 0.920 | -0.031 |
| fawiki | goodfaith | 0.961 | 0.897 | -0.064 |
| frwiki | reverted | 0.929 | 0.846 | -0.083 |
| hewiki | reverted | 0.874 | 0.800 | -0.074 |
| idwiki | reverted | 0.935 | 0.903 | -0.032 |
| itwiki | reverted | 0.905 | 0.850 | -0.055 |
| nlwiki | reverted | 0.933 | 0.831 | -0.102 |
| ptwiki | reverted | 0.894 | 0.812 | -0.082 |
| ptwiki | damaging | 0.913 | 0.848 | -0.065 |
| ptwiki | goodfaith | 0.923 | 0.863 | -0.060 |
| trwiki | reverted | 0.885 | 0.809 | -0.076 |
| trwiki | damaging | 0.892 | 0.798 | -0.094 |
| trwiki | goodfaith | 0.899 | 0.795 | -0.104 |
| viwiki | reverted | 0.905 | 0.841 | -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:
I've generated a dataset of editquality-modeling model scores for English Wikipedia. See http://datasets.wikimedia.org/public-datasets/enwiki/eq_studies/anon_scores.201612_week1.tsv
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
Looks like most of the signal comes from user features. See https://meta.wikimedia.org/wiki/Research:Building_automated_vandalism_detection_tool_for_Wikidata