Page MenuHomePhabricator

[Analytics] [Request] [WDPE] Wikidata Promotional Editing: Revert Risk Model - Item Revision Metrics Monitoring
Open, Needs TriagePublic8 Estimated Story Points

Description

Background:
Wikidata has always been a target for search engine optimization but it seems to be increasing significantly recently. People seem to be led to Wikidata as the place to anchor their entity to be accessible in generative AI answers etc. To help ensure promotional editing is being spotted and properly acted upon, a revert risk model has been created. The analytics team would like to support this work and provide metric monitoring for the Revert Risk Model usage and impact to this overall topic.

Questions & Metrics: Revert Risk Model

Question: Is the new risk model successful: are bad edits being reverted?

  • %/# of bad edits that get reverted.
  • % of bad edits that are reverted within 2 days.

Bad Edit Definition:
As per the work done in T429049, we can use the table event.mediawiki_page_revert_risk_prediction_change_v1 and the value from predicted_classification.probabilities['true'], below shows the ranges to choose for the model(s) and classifications of a bad edit:

TierDamaging (old)RR floor 0.641RR floor 0.371
likelygood range[<0.279][<0.650][<0.371]
likelybad range[0.385][0.650][0.371]
likelyverybad range[0.929][0.889][0.889]

Further example code here, in order to review and incorporate bot edits, correct revision IDs, etc.

Core source tables:

  • event.mediawiki_page_change_v1
  • event.mediawiki_page_revert_risk_prediction_change_v1
  • Join on page / revision IDs

**Deliverables**

  • Task code for bad edits and metrics
  • Snapshot of data exported
  • Superset dashboard
Data Snapshot Creation
  • Define metric(s) to be derived from the new / updated data pipeline
  • Define output table columns with stakeholder(s)
  • Create process diagrams in query documentation to map out complex flows
  • Write job queries or code
  • Test job queries or code
  • Export data snaphsot / add to wmde hdfs
  • Create Superset dashboard
Previous notes from initial session:

There are some notes / decisions to be made for the above metrics:

  # We need to firm up a definition of “bad edits”. See above added section for this definition based on T429049.
  # For these metrics we should include the option of filtering on item notability. Answer - we actually don't combine this as they are quite different.
  # Can we have an overview of training data? Not necessary, instead what is more useful are the findings from ticket T429049

Event Timeline

The definition of bad edits will come out of the thresholds the ML team is working on in T429049.

Team working on T429049 are actively working on the threshold calculations, so when it's completed can re-finalise the ticket

AndrewTavis_WMDE renamed this task from Wikidata Promotional Editing: Revert Risk Model to [Analytics] [Request] [WDPE] Wikidata Promotional Editing: Revert Risk Model.Aug 4 2026, 1:36 PM

Hi all - I've taken a look through the T429049 ticket, and asked a few clarifying questions, as well as whether they have monitoring set-up. But some further thoughts / questions for finalising this ticket scope:

  1. Searching for a "bad edit" definition: I'm wondering if we can simplify this slightly, as we have a table in the datalake mediawiki_page_revert_risk_wikidata_prediction_change_v1 where there is a field for the predicted classification of a page / edit. And if this value is TRUE then there is prediction that the edit should be reverted. Could we take this as our definition of "bad edit"? Anyone who has reviewed the commentary on T429049 will see though that there are still discussions / reviews for the validity of the classification. However, my understanding is that this has improved. And due to the nature of this work will never be perfect.
  2. If the above we are happy with, we then should confirm for this whether we are looking at edits or items (or both) for our metrics?
  3. I think the part of this task: Is the new risk model being used? is actually redundant. Due to the nature of the model roll out and who it's intended for. See here.

CC: @Lydia_Pintscher

Hi all - I've taken a look through the T429049 ticket, and asked a few clarifying questions, as well as whether they have monitoring set-up. But some further thoughts / questions for finalising this ticket scope:

  1. Searching for a "bad edit" definition: I'm wondering if we can simplify this slightly, as we have a table in the datalake mediawiki_page_revert_risk_wikidata_prediction_change_v1 where there is a field for the predicted classification of a page / edit. And if this value is TRUE then there is prediction that the edit should be reverted. Could we take this as our definition of "bad edit"? Anyone who has reviewed the commentary on T429049 will see though that there are still discussions / reviews for the validity of the classification. However, my understanding is that this has improved. And due to the nature of this work will never be perfect.

Like yes. Can you send me a sample so I can have a closer look?

  1. If the above we are happy with, we then should confirm for this whether we are looking at edits or items (or both) for our metrics?

For this we care about edits.

  1. I think the part of this task: Is the new risk model being used? is actually redundant. Due to the nature of the model roll out and who it's intended for. See here.

This was more about answering the question if the edits that the model classifies as bad are actually getting reverted by the editors or if they stick around for a significant amount of time. So I think we still want this. This will be especially important to monitor what's happening once Automoderator gets enabled.

Hey @Lydia_Pintscher

Like yes. Can you send me a sample so I can have a closer look?

Yes! Is this useful? Here you have for WD the page and revision IDs, when the probability > 0.641, which in the linked ticket is the classification for likely bad edit. Taken 100, from 23/07/2026

This was more about answering the question if the edits that the model classifies as bad are actually getting reverted by the editors or if they stick around for a significant amount of time. So I think we still want this. This will be especially important to monitor what's happening once Automoderator gets enabled.

OK, gotcha, then I think anyhow will need to re-write the suggested metrics in this ticket as they are not capturing this. It might make sense to put this into a separate task however.

Hey @Lydia_Pintscher

Like yes. Can you send me a sample so I can have a closer look?

Yes! Is this useful? Here you have for WD the page and revision IDs, when the probability > 0.641, which in the linked ticket is the classification for likely bad edit. Taken 100, from 23/07/2026

Great. The revision ID and probability is what we want I'd say.

catherine.kelsey.wmde renamed this task from [Analytics] [Request] [WDPE] Wikidata Promotional Editing: Revert Risk Model to [Analytics] [Request] [WDPE] Wikidata Promotional Editing: Revert Risk Model - Item Revision Metrics Monitoring.Wed, Aug 26, 12:39 PM

Fantastic @Lydia_Pintscher - as such, I've updated this task detail with some more info and split out the other metrics into a task here: T436091

Hi @Arian_Bozorg and @Lydia_Pintscher we just discussed in the office hours to make the deliverables of this ticket to be a Snapshot, as there still might be updates / tweaks being made to the revert risk model.

So we suggest, to run the metrics using the two different probability floors, for the available time period. Then, we can discuss how this can be followed up. Please let us know if this is not suitable.

Lena_WMDE set the point value for this task to 8.Tue, Sep 1, 2:37 PM