Page MenuHomePhabricator

Experiment: Topic-matched Event Invitations
Open, Needs TriagePublic

Description

Background

This is the second experiment in the Event Pathways initiative, following Test A (worklist-based matching). Where Test A triggers the post-edit invitation when an editor edits an article that appears directly on an event organizer's worklist, Test B explores a broader matching signal: topical overlap between the article being edited and an upcoming event.

Why a separate experiment for topic-based matching? Worklist matching is precise but limited in reach: it only fires when an organizer has explicitly added an article to their worklist. Topic-based matching could reach a much larger surface area of edits. This experiment tests whether topical matching can drive meaningful engagement, specifically among editors who are less likely to already know about events and more likely to find an invitation genuinely useful.

Testable hypothesis

If we invite editors to an event that matches the topic of the article they just edited, it will increase event registration rate because editors might not know about upcoming events that are aligned with their topical interests so inviting them to those should make them more likely register for an event.

Experiment details

Overview
  • Experiment Name: Topic-matched Event Invitations
  • Tracking Key: de-1-2-event-pathways-topic-matching
  • Risk Level: Tier 3: Low Risk
  • Contextual attributes: None
Traffic
  • Assign Variation by Attribute: MW User ID
  • Traffic Percent: 100%
  • Variations
    • Control (control): editors see no dialog after publishing an edit if their edit has a topical match
    • Invited (invited): editors see the post-edit event invitation dialog when their edit meets the trigger conditions above
  • Targeting
    • MediaWiki Database Code in:
      • TBD
  • Expected start date: TBD

Expected duration: The 2-week window is intended to capture a meaningful volume of qualifying edits while limiting exposure during an early test of topic-based matching. Will also depend on when the experiment reaches enough statistical power to ensure reliability of results. See https://wikitech.wikimedia.org/wiki/Test_Kitchen/GrowthBook_user_guide#Know_when_to_stop_the_experiment

Eligibility conditions

NOTE: To ensure that the Control and Invited groups are comparable, these eligibility conditions should be checked by the experiment's instrumentation for all enrolled users regardless of assignment.
  • The editor has fewer than 100 lifetime edits (newcomer or junior editor)
  • The wiki they edited on has an event starting within the next 14 days
  • The relevant event has a topical match to the article they just edited
  • The editor is not already registered for the matched event

Metrics

REMINDER: The linked specification at the top should list all these metrics and their events, and which (if any) contextual attributes are needed for specific events.

Notes

Out of scope:

  • Contribution rate – does the editor make at least one edit associated with the event via Collaborative Contributions?

Event Timeline

OBenhmida updated the task description. (Show Details)
mpopov renamed this task from Placeholder: Set up test kitchen for Event Pathways (part 2, by topic) to Experiment: Topic-matched Event Invitations.Jul 11 2026, 6:30 PM
mpopov added a project: Epic.
mpopov updated the task description. (Show Details)
mpopov added a subscriber: OBenhmida.

CHANGELOG

  • Task:
  • Metrics:
    • Event invitation clickthrough rate didn't make sense as a goal metric, since that's a feature-only metric so I made that a secondary metric (see guidance).
    • Event registration rate is the actual impact intended here.
    • For additional insights into why Event registration rate may or may not be impacted, we should also measure Median Event details page visits for both groups.
    • "Contribution rate – does the editor make at least one edit associated with the event via Collaborative Contributions?" was listed as a metric of tertiary interest. To simplify the experiment, I marked it as out of scope.
  • Experiment design:
    • Previously the control group included all editors, regardless of whether their edit matched the eligibility criteria. For a more accurate, appropriate comparison I have narrowed down the control group to have the same eligibility criteria as the Invited group, with the only difference between the two groups being that we actually show an invitation to the editor in the Invited group. This will greatly improve the trustworthiness of the experiment results.

(Per conversation in Slack https://wikimedia.slack.com/archives/C06M4A3SYMQ/p1783702580997159)

@MHorsey looked at how to match a just-edited article against the articles on an event's worklist, so we can show an event invitation on topical similarity rather than exact worklist membership only.

Two approaches were considered: CirrusSearch morelike (lexical similarity, one API call, works on every wiki, near real-time) and the LiftWing articletopic model (semantic topic labels, but natively covering around 10 wikis with weekly propagation elsewhere).

Decision: morelike for the prototype. It is simpler to implement, has no coverage gaps across wikis, and is good enough to validate the concept. articletopic remains a clear upgrade path if match quality proves insufficient, and can be layered in later without changing the architecture.

Two open items noted in the investigation: migrating SimilarArticleLookup from HTTP to SearchEngineFactory once the prototype is validated (a TODO marks this in the code), and whether the number of similar articles returned, currently 10, should be configurable per wiki given how much result relevance varies by wiki size.

Full write-up with the API testing and reasoning: https://docs.google.com/document/d/16nQ94T70sRI-P2LDjpi7_JznMF_6YjSW/edit?usp=sharing&rtpof=true&sd=true