The document discusses building a model to predict bias on Wikipedia pages. It outlines identifying patterns in editing behavior that typically lead to bias, like ownership behavior and opinion camps. These patterns will be used to train machine learning algorithms to predict bias. The goal is to help the declining number of editors address bias by developing tools that leverage the prediction model.

1. Towards a diversity-minded Wikipedia
Fabian Flöck, Denny Vrandecic, Elena
Simperl – KIT Karlsruhe

2. The loss of viewpoint diversity

3. Wikipedia is biased
o To what extend does Wikipedia represent all relevant
PoVs?
o Neutral-point-of-view policy
o Examples: Conservapedia, Croatian and Serbian editions

4. Leveraging diversity in Wikipedia
Use Case - In collaboration with Wikimedia Germany:
Displaying warnings when detecting patterns of bias in
editing behavior
Challenge:
Identify, understand and predict socio-technical
mechanisms leading to bias

5. Context: Slowing growth and content saturation
• Less new articles, less edit Total active editors per month
growth since 2007 (Thousands)
• Consolidation of articles
• No room for easy contribution
New articles per month
on en.wikipedia.org
Revert ratio per month
by editor class
Figures by Suh et al.
“The Singularity is not near: Slowing growth of Wikipedia”

6. Socio-technical mechanisms leading to bias
o Consensus: social proof and consolidation
o Ownership behavior
o Opinion camps and editor drop-out
o Boldness and useful conflicts
o Bureaucracy and implicit social norms
o Personal characteristics of authors
o WikiProjects and other groups

7. Building a bias prediction model
1. Identify typical patterns for a socio-technical mechanisms
2. Evaluate if these patterns are typical for articles marked as
“biased”
3. Train machine learning algorithms to predict bias using the
patterns

8. Building a bias prediction model
1. Finding patterns:
o Ownership behavior
• Typical patterns in “Maintained” articles
 High # of reverted newcomers
 High concentration of edits, reverts
 Core high-activity-editor group
 …
• Methods
 Statistical analysis on revision history, etc.
 SNA
o .. And for all the listed mechanisms

9. Conclusion
o Building a prediction model as complete as possible
o Converge with other models developed for Wikipedia in
RENDER
 Build tools that help the shrinking number of editors to cope
with the work overload of finding and curing bias