Research:Topical coverage of Edit Wars

From Meta, a Wikimedia project coordination wiki
Jump to: navigation, search
Created
23:06, 16 September 2017 (UTC)
Duration:  2017-September — 2017-December
GearRotate.svg

This page documents a research project in progress.
Information may be incomplete and change as the project progresses.


Introduction[edit]

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 [1].

Methodology[edit]

  • 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.

Topic Model[edit]

Pages to Topic[edit]


Topic Distance[edit]

  • 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 [2] 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.

User Behavior[edit]

Topical Coverage[edit]

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 (upT : 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 psports is 3/T = 7.3e-08.
  • Topical stability (us): 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.

Reverting Behavior[edit]

  • 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.

Probability of Reverting Other User[edit]

  • TODO

Dataset, Tools and Resources[edit]

Revisions[edit]


Wikiprojects Graph[edit]

Code[edit]

  • Find all the code used in this study here:

Terminology / Conventions[edit]

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

Results[edit]

(Work in progress... )

Characterization on Topical Distance in Multipage editwars[edit]

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.

Characterization Topical Distance in Multipage Editwars
Distance  %
-1 0.7107
0 0.2297
4 0.0143
5 0.0121
2 0.0103
3 0.0074
6 0.0056
7 0.0034
8 0.0027
1 0.0016
9 0.0012
10 0.0007
11 0.0002
12 0.0001

Future Work[edit]

  • 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)

Timeline[edit]

Q1, Q2

References[edit]

  1. Precision_and_recall
  2. shortest path
  3. script for parsing wikiprojects graph

Acknowledgements[edit]