Research:Topical coverage of Edit Wars: Difference between revisions

This page documents a research project in progress.
Information may be incomplete and change as the project progresses.
Please contact the project lead before formally citing or reusing results from this page.

This project is run by the [[mw:Wikimedia Research|Research Team] as part of the Community health initiative. Other (parallel) research studies focus on the usage of toxic language ^[1] or specific behavior such as wiki-hounding ^[2], mainly working with the content of user interactions. Here we adopt a complementary language agnostic approach, that focus on edit wars, trying to differentiate topic-centered from person-centered conflicts.

Edit wars in Wikipedia has been largely studied. An edit war is usually consider to be the consequence of different opinions about an specific topic between two or more users. Naturally, users with different political views might have different opinions on many articles related with politics, and these differences can scale in a multi-article edit war. These actions can be consider toxic, but is not necessarily a stalking behavior. However, in the case (if exist) that edit wars start happening across multiple topics, this can be an indicator of a person-centered attack (instead of topic-centered), that might be categorized as wikihounding.

Taking the advantage that edit wars can be detected in content agnostic approach (without analyzing the text), we propose to study the topical span of those wars, characterizing usual and unusual (potentially toxic) behaviors.

The main tasks to develop such model are:

• Generate a representative dataset of edit war in Wikipedia.
• Detect pairs or groups of users involved in more than X (define X part of the study) controversies.
• Define and implement a robust topic model.
  • Define a distance metric for topics (eg: Geography is N steps far from Politics, and M steps far from Sports.)
• Apply an outlier detection mechanism to find potential cases of harassment .
• Our approach focus in obtain good recall ^[3].

• Define a topic model, that allows to measure topical distance between Wikipedia pages.
• Model user behavior according to the topics that she/he covers (edit on) and the amount of reverts that she/he does.
• Compute the probability of pair of users of co-edit a page, and the probability that this co-edition is a revert.
• Based on the aforementioned co-editions' probability, identify anomalous behavior that are potentially related with stalking or wikihounding behavior, and do manual evaluation.

• We use wikiprojects as proxy for topic assignment.
• Mapping from pages to Wikiprojects can be obtained with this query: https://quarry.wmflabs.org/query/23214
• Therefore, we consider each wikiproject as one topic, however, usually each page belongs two more than one topic.
• Each page can belong to 0, 1 or several Wikiprojects. For example the page about en:Diego_Maradona is the interest of en:Wikipedia:WikiProject_Biography/Sports_and_games, en:Wikipedia:WikiProject_Football and en:Wikipedia:WikiProject_Argentina

• Wikiprojects can be represented as a graph (we use this library https://github.com/wiki-ai/drafttopic/tree/master/drafttopic/utilities).
• Then, giving that each page can belong to more than one category, we define the distance between two pages as the minimum shortest path ^[4] among all pairs of nodes on the Wikiprojects graph.
  • Example: Given a page X and Y, with X in Wikiprojects a and b and Y in Wikprojects c, and d. We compute the lenght of shortest path between (a,c), (a,d), (b,c) and (b,d), and return the minimum value among these results.

We use two metrics to characterize the topical focus/coverage of a given user, one focused on the probability of editing on a giving topic, and other focused on the user topical stability:

• Probability of U on editing on topic T (u_pT) : The probability of user U editing topic T, is the her total number of editions on T divided by total number of editions in the set. Example: in our dataset T= 40,833,278 (total number of revisions), consider a user U, with 3 editions on topic Sports, his probability of editing on topic sports p_sports is 3/T = 7.3e-08.

• Topical stability (u_s): For each user U, we obtain the distance for his next edition . For example, given a user U, doing three editions, the first one in the topic 'Sports', the second one in the same topic, and the third one in the topic biology (with distance 4 from Sports), the probability of user U to edit with topical distance 0, is 2/3, with distance 4 is 1/3, and 0 for the other distances. This metric gives an idea of user stability in terms of topics.

• For user we compute portion of reverts (within the dataset) compared with his/her total amount editions. Considering that our dataset just contains editions done by users 10 or more editions, we are just considering the reverts among these users. Later, in the results section we discuss the implications of such constrain.

• TODO

• We consider all the users with more than 10 revisions in the selected period (1/Jan/2017 to 16/Nov/2017)
• All the revisions from those users can be obtained from this query: https://quarry.wmflabs.org/query/23424
• Pages to Wikiprojects: https://quarry.wmflabs.org/query/23214

• Using the Wikiproject graph generated from this script^[5]

• Find all the code used in this study here:

• We use edition as synonym of a Wikipedia revision.
• We use Wikiproject as proxy for Topics

(Work in progress... )

In order to have notion of how frequent are reverts and editwars across multiple topics, we consider all pair of users with U, V, where U has reverted V more than 2 times, and compute the topical distance between all the pairs of pages reverted, next we compute the mode (most frequent value) for each pairs of users, and report the frequency of those values. As expected, most of wars (71%) focus on one page, 22% in the same topic but different page. The remaining 7% are a cross-topic reverts, reinforcing our intuition that cross-topic edir wars are rare.

• Improve mapping system from pages to Wikiprojects: The aforementioned [query https://quarry.wmflabs.org/query/23214] returns around 22% of pages matching with no wikiproject. However, manually reviewing we found cases where the Wikiproject is not correctly (?) assigned as category in the Talk pages (example en:Classon_Avenue_(IND_Crosstown_Line) belongs to Wikiproject Trains, but that project is not listed as category)

Q1, Q2

• User:Halfak (WMF)
• User:Sumit.iitp
• User:Jmorgan (WMF)