Page MenuHomePhabricator

Report on baseline metrics for recommendations experiments (UI and non-UI experiment)
Closed, ResolvedPublicSpike

Description

Background

  • We will be testing various APIs that provide recommendations against one another. As part of this we would like to measure how frequently current recommendations are used

User story

  • As a product team, we want to know how the current Related Changes and other recommendation systems perform

Requirements

  • Explore the data and report on the following for the UI experiment baselines:
    • Daily clickthrough rate for the Related Pages feature on mobile
    • Daily overall percentage of internal referrals generated by the Related Pages feature on mobile

Event Timeline

ovasileva triaged this task as High priority.

Jennifer and Olga were planning to meet yesterday (Monday 26th). Anything to update on this ticket here?
Clarity on this ticket would help us unblock work on T372780.

ovasileva renamed this task from Identify baseline metrics for non-UI recommendations experiment to Identify baseline metrics for recommendations experiment.Sep 2 2024, 10:39 AM
ovasileva reassigned this task from ovasileva to jwang.
ovasileva updated the task description. (Show Details)

Jennifer and Olga were planning to meet yesterday (Monday 26th). Anything to update on this ticket here?
Clarity on this ticket would help us unblock work on T372780.

@Jdlrobson - just to clarify, this ticket is for gathering data only, specifically for measuring the current usage of related pages. The relevant ticket for metrics for the search UI recommendations experiment (the browser extension) is T371236: Identify metrics for search UI recommendations experiment

ovasileva renamed this task from Identify baseline metrics for recommendations experiment to Identify baseline metrics for recommendations experiments (UI and non-UI experiment).Sep 2 2024, 10:42 AM
ovasileva renamed this task from Identify baseline metrics for recommendations experiments (UI and non-UI experiment) to Report on baseline metrics for recommendations experiments (UI and non-UI experiment).
ovasileva updated the task description. (Show Details)

The Search platform team does track Related Pages in the following graph
Related Articles pageview actors as portion of all pageview actors; Plus clicks-to-impressions

These metrics are based on queries to the discovery.webrequest table, which tracks both the number of "morelike" API requests and pageviews that include the &wprov=rarw1 parameter. The wprov parameter is added to each Related Articles link. While this metric does not directly track actual click events, it does reflect the outcome of a click on a Related Articles link.

@Jdrewniak , what is morelike API request? what action triggers such a request? For context, I am looking for the "impression" events to estimate the click through rate of Related Articles link.

hi @jwang the "morelike API request" uses the webrequest table to track requests to an API like this:
https://en.wikipedia.org/w/api.php?action=query&list=search&srsearch=morelike:radium&srlimit=3&srprop=size&formatversion=2

Currently, this API is only used in the RelatedArticles extension to show related articles at bottom of the page, on mobile (screenshot below), so essentially, this metric seeks to track the number of relatedArticles impressions (each impression is one API request).

Screenshot 2024-09-09 at 6.13.17 PM.png (892×2 px, 245 KB)

Daily clickthrough rate for the Related Pages feature on mobile

  • Based on the data between 2024-08-09 and 2024-09-08: around 5.96% of Related Articles presentations result in a click.
MetricTotal (Sum)
Impressions2.6B
Pageviews of related pages ( equivalent to clicks on Related Article links)155M
Click though rate5.96%
  • Daily trend

image.png (678×1 px, 63 KB)

Note: The clickthrough rate is lower than reality. As the users might not actually see the related page links. More details are documented in superset dashboard presented by search platform team.

Daily overall percentage of internal referrals generated by the Related Pages feature on mobile

  • Based on the data between 2024-08-09 and 2024-09-08: around 4.20% of internal pageviews are generated by the Related Pages feature on mobile
MetricTotal (Sum)
Internal Pageviews3.69B
Related Articles Pageviews155M
Percentage4.20%
  • Daily trend

image.png (1×1 px, 84 KB)

Note: there was a dip between 8/7 and 8/16 due to a decrease in the related article pageviews.
Superset link: https://superset.wikimedia.org/superset/explore/p/zWGn1KyXV9y/

Based on the discussion with @Jdlrobson on slack, it would be helpful to have the daily clickthrough rate for the Related Pages feature on mobile and desktop on English Wikivoyage. Here is the data.

Daily trend of CTR on English Wikivoyage

image.png (1×1 px, 96 KB)

Summary:
On English Wikivoyage, mobile web has a higher click through rate (CTR).
Daily average CTR on desktop on English Wikivoyage: 6.34%
Daily average CTR on mobile web on English Wikivoyage: 9.93%

Superset link: https://superset.wikimedia.org/superset/explore/p/OjRbj9mXNZ0/

Thank you @jwang! @Jdlrobson, @jwang - curious about how we're measuring impressions here. Which of the following is an impression in this context:

  1. All pageviews
  2. All times when related pages is shown (measured by whether someone scrolls to the bottom of the page?)
  3. Something else?

@Jdlrobson - also curious why we looked at wikivoyage specifically - is this related to our upcoming experiment in any way?

@Jdlrobson - also curious why we looked at wikivoyage specifically - is this related to our upcoming experiment in any way?

I suggested this as Wikivoyage is the only project where RelatedArticles is enabled on desktop and I felt this data point might be useful given our experiments are on desktop and this analysis on Wikipedia's is mobile only.

I suggested this as Wikivoyage is the only project where RelatedArticles is enabled on desktop and I felt this data point might be useful given our experiments are on desktop and this analysis on Wikipedia's is mobile only.

Got it. That's a really good call!

In terms of impressions - from standup discussion, we count an impression as the MoreLike API being called which happens when a user is near the related articles feature but we cannot confirm 100% that the feature appears on the screen when the call is made.