Us vs.

Them: Understanding Social Dynamics in Wikipedia with Revert

Graph Visualizations

Bongwon Suh, Ed H. Chi, Bryan A. Pendleton, Aniket Kittur

Palo Alto Research Center

ABSTRACT repositories on the Internet. A recent study comparing it against

Wikipedia is a wiki-based encyclopedia that has become one of the Encyclopedia Britannica reports that much of the Wikipedia
content is of high quality in spite of issues such as vandalism and
the most popular collaborative on-line knowledge systems. As in
factual inconsistencies [14].
any large collaborative system, as Wikipedia has grown, conflicts
Since its inception, Wikipedia has been growing at an
and coordination costs have increased dramatically. Visual
exponential rate [5, 6, 18, 32, 47]. Over two million articles have
analytic tools provide a mechanism for addressing these issues by been collaboratively edited by more than four million users in the
enabling users to more quickly and effectively make sense of the English Wikipedia alone. However, the high level of participation
status of a collaborative environment. In this paper we describe a comes with corresponding costs, including conflict between user
model for identifying patterns of conflicts in Wikipedia articles. factions, vandalism, spam, and expression of political ideologies.
The model relies on users’ editing history and the relationships Keeping any social structure or organization growing often
between user edits, especially revisions that void previous edits, requires overhead costs, such as coordination and maintenance
known as “reverts”. Based on this model, we constructed Revert work [24, p.160]. Wikipedia is not an exception and our earlier
Graph, a tool that visualizes the overall conflict patterns between work showed that the overhead cost in Wikipedia has increased
groups of users. It enables visual analysis of opinion groups and significantly as Wikipedia evolves over time [18]. We showed
rapid interactive exploration of those relationships via detail drill- that some of these coordination costs result from user
downs. We present user patterns and case studies that show the disagreements about article content, procedures, and
effectiveness of these techniques, and discuss how they could administrative issues.
generalize to other systems. Visual analytics may provide a mechanism for users to reduce
the costs of making sense of the state of collaboration and
CR Categories and Subject Descriptors: H.5.3 [Information engaging in the above activities. The result of the visual analysis
Interfaces]: Group and Organization Interfaces – Collaborative could help users find major opinion groups and subjects of
computing, Computer-supported cooperative work, Web-based controversy, identify mediators and abusive editors, and
interaction understand the user conflicts within these systems.
Additional Keywords: Wikipedia, wiki, revert, graph,
collaboration, user model, visualization

1 INTRODUCTION
Human existence depends on collaborative problem solving.
Nations and companies depend on collaboration to solve vexing
problems such as rules of conduct, investment decisions, and
resource allocation. It is therefore not surprising that visual
analytics must work effectively in collaborative environments.
Indeed, diverse fields that utilize visual analytics such as scientific
inquiries into climate change and foreign intelligence analysis on
Iraq insurgents rely on collaborative problem solving to finally
reach actionable conclusions. These conclusions depend on
groups of people forming a coherent picture of the problems at
hand, and then developing a consensus amidst conflicting user
opinions and political pressures. Recently, new collaborative
knowledge systems such as Wikis are being used for collaborative
problem solving and knowledge gathering [2]. As conflict and
coordination costs increase in such environments, visual analytic
tools may be increasingly useful for users to make sense of the
status of the collaborative environment.
The largest experiment of this kind is probably Wikipedia,
which has become one of the most popular knowledge

Address: 3333 Coyote Hill Road, Palo Alto, CA 94304

Email: {suh, echi, bp}@parc.com, nkittur@ucla.edu
Figure 1. Revert Graph visualizes social structures in Wikipedia
articles. Revert Graph for the Charles Darwin page [37]. For a
discussion, see Section 6.
IEEE Symposium on Visual Analytics Science and Technology 2007
October 30 - November 1, Sacramento, CA, USA
978-1-4244-1659-2/07/$25.00 ©2007 IEEE

163
 In this paper, we develop a user conflict model based on users’ such as McGrath to examine theories of how groups collaborate
editing histories, specifically revisions that void previous edits, over time, and how they resolve conflicts [19].
known as “reverts”. Our model extracts reverts from Wikipedia Indeed, communities of practice have changed greatly since the
editing history and composes a node-link graph where a user is introduction of the Web and online communities [3, 7, 9].
denoted as a node and a revert relationship as a link. Based on this Arguello et al. studied how individuals must act in order to
model, we developed a tool called Revert Graph that visualizes receive a reply in a discussion forum [3]. Cosley et al. studied
the revert relationships between opinion groups. It utilizes a how intelligent task routing can help with individuals wishing to
force-directed layout to cluster user groups, and provides detailed collaborate and create useful value to social groups [7]. Dibbell
drill-down to help identify specific user opinions. described how conflicts, vandalism, and anti-social behavior
We shall show that the tool can help discover and pinpoint user manifest itself in virtual worlds such as MUDs and MOOs [9].
patterns such as the: (a) formation of opinion groups; (b) patterns Dourish and Bellotti studied how awareness is required for
of mediation; (c) fighting of vandalism; (d) identification of major coordination in shared workspaces [10].
controversial users and topics. The tool can be used to identify Viewed from this CSCW perspective, the rise of conflict and
the severity and nature of a disagreement and the number and the costs of coordination are unavoidable in a distributed
composition of the user groups involved. Figure 1 shows some collaboration system [13] such as Wikipedia, and manifest in
example social structures discovered and characterized using scenarios such as conflicts between users, communication costs
Revert Graph. We believe the tool can form the basis for conflict between users, and the development of procedures and rules for
resolution tools in the future. coordination and resolution. Researchers have seen similar costs
The contributions of this paper include: in other computer mediated communication (CMC) systems [8,
• the development of a user conflict model for wiki-style online 9]. Even though researchers have documented the growth of
knowledge repositories; Wikipedia [5, 32], the impact of coordination costs for adding
• the creation of a visualization tool that enable analysts to content and users has not been well studied, though Buriol et al.
understand and explore user conflicts and opinion groups; [5] discussed maintenance cost such as reverts and vandalism.
• the presentation of user patterns and case studies that show the Perhaps the most relevant study to coordination and conflict
effectiveness of these techniques. cost is our previous study [18]. In the study, we discovered that
The rest of this paper is structured as follows: First, we present the number of article edits is decreasing while the overhead (i.e.,
related work on Wikipedia social dynamics, collaborative problem the number of edit intended for communication and policy
solving, conflict resolution, and visualization of social making) is increasing. The study also introduced a model for the
environments. Second, we introduce Wikipedia and its particular degree of conflict based on simple metrics. According to the
structure and statistics. Next, we introduce the user conflict study, the increased number of revisions made on article
model as well as the particular visualization method we used. We discussion pages is the most significant predictor for the degree of
then demonstrate the tool by presenting interesting user patterns conflict in an article. This finding strongly hints that the degree of
and case studies we have found in Wikipedia. We conclude by conflict and disagreement between users has a close relationship
discussing how the tool can generalize to other systems, future with the amount of total overhead. Motivated by earlier results,
directions, and final remarks. here we focus on developing a user conflict model for Wikipedia.
Visual analytics [29] offers a potential solution for
2 RELATED WORK understanding conflicts and coordination costs in collaborative
Collaborative knowledge spaces and collaborative problem social spaces. For example, researchers have visualized conflicts
solving are big research areas with many researchers working on between political blogs [1] and entities in political news [4].
different aspects of the problem. In order for these collaboration Researchers have also used visualization to understand social
spaces to be effective, we need to understand how conflicts arise spaces by visualizing relationships amongst people in chat rooms
and how conflicts are resolved in these systems. Here we [30], friends [17], and other social networks [23].
summarize some past works in social dynamics and conflict Visualization also offers a potential way to understand the
resolution. dynamics of content creation between users of a collaborative
There appears to be a number of intuitions about why space. Viegas et al. showed how visualization can be used to
collaboration might improve analytics [24, Chapter 8]. First, an understand the editing history of a single article in Wikipedia
individual receiving information cues from a group of other [31]. The tool was able to reveal some social patterns such as
analysts is more likely to be more efficient in exploring a domain vandalism, negotiations, edit wars, and anonymous edits.
area. Second, since the work covering a large domain area can be Visual analytic tools not only can help us characterize the
divided up into the light work of many, one can ensure important collaboration space such as Wikipedia, but they also potentially
pieces of evidence are not missed. Third, knowledge could offer benefits to end-users who need to understand how
specialization ensures that years of skilled training in one others in their social groups are performing, what they are paying
particular area might enable one to build up better ability to spot attention to, and what conflict patterns might exist during
patterns. Fourth, the diversity of viewpoints can potentially collaboration. In other words, visual analytics applied to these
overcome cognitive biases toward one single interpretation of the social spaces will enable a kind of situation awareness [11] of the
data. analytical topic area. These tools reveal social structures and user
Researchers have studied this area in the field of Computer behavior that shed light on how we can characterize conflicts and
Supported Cooperative Work (CSCW) and tried to understand coordination costs.
distributed problem solving and cognition. Documents and In this paper, we employ a novel method for constructing the
spreadsheets often are focuses of collaboration. For example, conflict model between users in Wikipedia, and apply visual
various researchers have studied collaborative writing and analytic techniques to understand the social relationships between
commenting [21, 22]. Nardi and Miller found that spreadsheets these users. We show how the tool is able to find interesting
are used as collaborative tools for problem solving in patterns such as the formation of opinion groups and mediation
organizations [20]. Flor and Hutchins described how distributed patterns, and controversial editors.
cognition occurs in a team during software development [12].
Changes in social computing spaces have caused psychologists

164
 Our recent study [18] found that work going into editing article
3 INTRODUCING WIKIPEDIA pages has been steadily decreasing from 95% to 65% (see Figure
Before we explain how we have identified conflict between users 2). Meanwhile, work going into communication purposes has
in Wikipedia, we first introduce relevant Wikipedia structures and been increasing to about 15–20% of total edits. Vandalism has
policies, and describe the data that we used in the analysis. also been increasing but still accounts for less than 1% of edits.
100%
3.1 WIKIPEDIA DATA ANALYZED Maintenance
95%
In this paper, we used a complete history dump of the English
Wikipedia that was generated on July 2, 2006. The dump included 90%

Percentage of total edits

over 58 million revisions, of which 2.4 million are article entries 85%
Other
in the encyclopedia, totaling approximately 800 gigabytes of data.
To process this data, we imported the raw text into the Hadoop 80%
User Talk
[15] distributing computing environment running on a cluster of 75%
commodity machines. The Hadoop infrastructure allowed us to User
70%
explore new full-scale content analysis techniques quickly. We Article Talk
also cloned Wikipedia’s own metadata databases for direct 65%
analysis. This database allowed us to inspect Wikipedia statistics Article
60%
in their native format.
2001 2002 2003 2004 2005 2006

3.2 WIKIPEDIA ORGANIZATION

Figure 2. Changing number of edits over time made on various
Here we briefly introduce the internal organization of Wikipedia: types of pages in Wikipedia
Pages: To manage the work on the encyclopedia while still
presenting a usable encyclopedia interface to non-participants, Recent research shows that Wikipedians spend a significant
Wikipedia organizes its editable pages into several categories. The amount of their effort on coordination and conflict resolution [18].
encyclopedia content itself is kept on article pages, while This strongly indicates that there exist underlying social processes
discussion of ongoing changes and other disputes to a specific that govern coordination. In this paper, we focus on exploring the
article are kept on an associated discussion, or “talk” page. social dynamics, which is crucial to understanding the
Similarly, individual contributors have a unique user page, and an coordination and conflict resolution process in Wikipedia.
associated user “talk” page to enable discussion with or about a
user. There is also a revision comment field, which is a free-form 4 CONFLICT USER MODEL
text field supplied by each user when submitting an edit, intended To understand disagreement between users, we need to build a
to describe what kind of changes were made. model of how users engage in disputes. One way to discover
Policies: Most activities on Wikipedia are governed by various disagreements is to look in the revision log and analyze the
policies. Users are encouraged to follow a general policy to Be substance of each user’s edits. However, we found this approach
Bold [36], i.e., take action themselves when there is an apparent rather impractical for large data sets such as Wikipedia, because it
reason. Another policy is to refrain from altering content so that it requires accurate identification of the intention of each edit, which
presents a non-neutral point of view [41]. A further policy asserts is not viable for machine-only processing. Many edits may have
that material should be attributed to a reliable published source non-obvious intentions based on the context in which it was made.
[35]. Together, these and other policies are also subject to the Instead, we use users’ editing history to approximate
same editing and review process that governs the encyclopedia disagreements between users. Specifically, we use revert revisions
content, but play an intrinsic character in determining the social as a way of identifying disagreement and conflicts between users.
norms that the communities is expected to adopt. The user survey from the previous study enabled us to collect
Users: Users of Wikipedia fall into distinct categories. Since it insight that Wikipedians often treat being reverted as an indication
is not required for a user to register to edit most pages, many users that another user strongly disagrees with an edit [18]. We employ
are anonymous. Users may establish a user account, which allows revert as a proxy for dispute and develop a user model based on
them to maintain a consistent identity across different sessions revert activity, described in the next section.
and network addresses. Additional permissions can be granted to
such users. The most significant of which is the role of 4.1 REVERTS IN WIKIPEDIA
Administrator, who has the ability to change many settings, such A revert is defined as undoing the actions of another editor in
as locking a page for editing, banning abusive users, and the whole or in part [44]. Reverts are often used to fight vandalism to
ability to access content which has been otherwise deleted from bring articles back to their original state. However, users also use
the encyclopedia. reverts to block other users’ contributions. Edit wars [40] are a
typical example, where disagreeing users repeatedly revert each
3.3 COORDINATION AND MAINTENANCE COST IN WIKIPEDIA others’ edits. Revert is a very reliable sign that one user clearly
Users interacting with each other create conflicts and incur opposes other’s edit, and, therefore, is a reasonable proxy for
coordination costs. These interactions take place on user and underlying conflict and disagreement.
discussion pages, which constitute a category of work we label We identify reverts in Wikipedia by two different methods. In
“overhead”. Overhead work represents the coordination costs of the data driven method, we computed a unique identifier of every
working in such collaboration spaces. One form of overhead revision made to every article using the MD5 hashing scheme
consists of edits made to non-article pages. These overhead [25], which is commonly used to check that data objects are
activities can be roughly divided into three categories: (1) identical. We use the hashing function to generate a small
maintenance work such as formatting and fighting vandalism, (2) fingerprint of each revision, which is suitable for rapidly
developing administrative procedures and policies for Wikipedia, comparing all revisions of an article. Using MD5 values for all
and (3) communication cost between users such as edits on article revisions of an article, we can identify when a later revision
talk, user talk pages.

165
 exactly matched the hash of a previous article, indicating a revert. • Identity Based Revert: the MD5 method is used to identify
The advantage of this method is that it does not depend on users reverts. When two revisions have the same textual content, we
to label reverts, which is not always done consistently. define the later edit as revert.
However, the disadvantage of this method is that it does not • Immediate Revert Only: When an article page is reverted to an
pick up partial reverts, in which only some of the text in an article older version other than its immediate last version, the intention
is reverted. To capture partial reverts we used a user-labeled of the revert is ambiguous because it is not clear whether the
metric, counting revisions whose revision comments included the revert is exclusively toward the last edit. We only count the
text “revert” or “rv” (a commonly used abbreviation of revert). revert relationship between the reverter and the editor who
The combination of both the data-driven and user-labeled methods made the immediate last edit regardless of this ambiguity.
provides converging evidence on the true change in reverts over
We employ these principles to build our visual analytics tool
time.
called Revert Graph.
Table 1 shows that the statistics for reverts calculated by the
two methods have slightly different characteristics. The MD5 5 REVERT GRAPH – VISUALIZING USER CONFLICT
identity revert discovery technique captures more revisions than
user-labeled (comment) reverts (3.7M vs. 2.4M), suggesting that a Revert Graph is designed to visualize social relationships between
substantial number of reverts are not labeled by users as such. Wikipedians as a node-link graph. The tool not only visualizes the
The union of the two methods provides the most accurate view of conflict model generated from editing history, it also provides
reverts, resulting in 3,917,008 reverts marked by either comments useful functions to investigate interactions between users.
or MD5 hashes. In other words, approximately 6.7% of all
changes in Wikipedia goes to restoring articles to previous 5.1 LAYOUT PRINCIPLES
versions. Revert Graph is a tool that enables an analyst to quickly
Vandalism in Table 1 is calculated by a similar method to the understand the relationships and patterns of activity that embody
comment method for reverts. We looked through the revision the conflict between users. Our layout algorithm simulates the
comments of each article for any form of the word “vandal” or social dynamics that result from the user conflict model. We
“rvv” (“revert due to vandalism”), which are put there by users accomplish this by visually gathering users with similar or
when removing vandalism. We showed that vandalism appears to compatible opinions together, while separating disagreeing users.
be increasing as a proportion of all edits, though it remains at a We implemented this approach using a force-directed graph
fairly low level (1-2% of all edits) [18]. layout algorithm [16] that assigns forces such that the edges
(representing revert relationships) act as springs, while the
Users Total 3,769,347 individual users are represented as particles with gravitational
Users who made at least one revert 402,454 fields (as shown in Figure 3).
Revisions Total 58,545,791

Reverts (MD5 hash method) 3,711,638

Self-reverts 582,373
Pages with at least one revert 721,866
Pages with 50 reverts or more 9,973
Reverts (Comment method) 2,422,482
Vandalism (Comment with vandal, rvv, etc) 577,643
Reverts (Union of both methods) 3,917,008
Table 1. User, Revision, Revert and Vandalism Statistics Figure 3. Force directed layout structure employed in Revert Graph.
Users (represented as nodes) attract each other unless they have a
4.2 REVERT-BASED USER CONFLICT MODEL revert relationship. A revert relationship is represented as an edge,
Our user model is based on the revert history between users thus pushing such users apart. Left figure: Nodes are evenly
distributed as an initial layout. Right figure: When forces are
assuming reverts are proxies for dispute and disagreement.
deployed, nodes are rearranged in two user groups.
However, we found a number of issues when applying reverts to
build a conflict model. Initially, the tool loads a group of users participating in editing
Using reverts to identify conflicts is hard because 1) multiple an article as a uniformly distributed node-link graph. As the
users are often involved in chains of reverts; 2) edit history is simulation runs, forces in the graph stabilize, and social structures
typically long and tedious to browse; and 3) there exist various between users begin to emerge, as shown in Figure 4.
types of reverts such as the “revert duel” (repeated reverts Node size is proportional to the log of the number of reverts or
between two users), self-reverts, reverts by multiple-account users revisions. Nodes are color-coded based on users’ registration
(or sock puppets), and so on. To address these challenges, we status. An administrator is drawn as a green node, a normal
made a number of design choices for our conflict model: registered user as a grey node, and an unregistered anonymous
• Disregard Self Revert: Self-reverts are disregarded. user as a white node, respectively.
• Degree of Conflict: the amount of dispute between two users is
measured by the number of reverts between them.
• Conflict Group: When two users make reverts on edits made by
another user, but not against each other, the two users are
presumed to have similar opinions. We will discuss this
principle in more detail in the following section.

166
 As social structures emerge from the force-directed layout, the
Group D tool provides users to drill down the graph allowing investigation
to the level of an individual revert (Figure 5). When a user node
is chosen in the graph, the upper right window displays the list of
Users having revert
Group A users that have revert relationships with the selected user, sorted
relationship with the
by the number of reverts between them (Figure 5 (b)). When a
selected editor
revert relationship is selected in this list, the bottom right panel is
updated to show individual revert records between the two users
involved in the revert revision, as shown in Figure 5 (c). Also, the
nodes representing the users are highlighted in Revert Graph to
provide visual feedback. Allowing further drill down, clicking an
item in the bottom right window launches a web browser showing
the specific individual revert record.
Group B Individual revert Using these capabilities, we were able to identify a number of
Group C history between the interesting user conflict patterns using this tool. In the next
two selected editors section, we will describe these conflict patterns in several case
studies.
Figure 4. Revert Graph for the Wikipedia page on Dokdo [39] 6 USER CONFLICT PATTERNS AND CASE STUDIES
Revert Graph uses force directed layout to simulate social
Based on the revert-based user conflict model, we investigated
structures between users. The tool also allows users to drill down
conflicts and disagreements in Wikipedia using Revert Graph and
into an individual revert revision enabling detailed investigation
explored to discover social patterns in them. This section presents
about the nature of the conflicts. Group A: mostly users with
interesting conflict patterns we have found in Wikipedia.
Korean point of view, Group B: mostly users with Japanese point
of view, Group C: mixed point of view, Group D: primarily non-
6.1 METHODOLOGY
registered users.
We used Revert Graph to examine conflict patterns in Wikipedia
5.2 USER INTERFACE articles. We selected 901 high conflict articles with more than 250
The majority of the screen is devoted to three windows: Revert reverts for analysis. These articles contain a large amount of
Graph itself is in the center; a list of revert relationships for the discussion with extensive editing history, which present a
selected user is on the upper right; and the selected individual challenge for analysts in making sense of conflict dynamics [18].
revert history is on the bottom right (Figure 4). Revert Graph also Based on the user model, Revert Graph generates a node-link
provides ways to change zooming level, node size, and other diagram to visualize users and their revert relationships. We then
visual options. Revert Graph is designed to help identify user analyzed the graph for any interesting pattern that might emerge.
groups representing opinion groups, the specific motivation of To examine a potential user conflict pattern in an article, the
analysis involved detail investigation of the article revision
revisions, and the conflict detail.
history. However, we often found it hard to determine users’ point
Suppose a user wants to investigate conflicts and disagreements
of view by browsing only the revisions. To get more clear insight
inside a Wikipedia article. The tool allows the user to specify an
on users’ position on the issues of an article, we browsed through
article she wants to explore by typing the name of the article.
information such as revert comments, article talk pages, user
Then the revert history of the article is retrieved from our database
pages, and users’ edits on other pages.
and a node-link graph is formed and displayed on the screen. A
We now describe user conflict patterns we found in this study.
force-directed layout module then clusters user nodes based on
revert relationships. Figure 4 shows an example. 6.2 PATTERN ONE – NODE CLUSTERS AND OPINION GROUPS
Revert Graph rearranges its user nodes based on revert
relationships between them. The force-directed layout simulation
evolves the graph to gather user nodes together based on
underlying social dynamics. We analyzed node clusters to
understand cohesiveness in node groups.
The Wikipedia page on Dokdo is one example where we were
able to find interesting user groups. Dokdo is a disputed islet in
(b) Drill down to list of revert the Sea of Japan (East Sea) currently controlled by South Korea,
relationships Alienus is involved in but also claimed by Japan as “Takeshima” [39]. Figure 4 shows
opinion groups discovered on the Dokdo article. We manually
labeled users based on their points of view as exhibited by their
editing history. To obtain users’ points of view on the topic, we
(a) A user node, Alienus is browsed their user pages, user talk pages, revision histories,
selected in Revert Graph (c) Drill down to revert history revision comments, as well as specific reverts. For example, users
between two users in group A in Figure 4 exhibit the following patterns: (1) claiming
Korean heritage on their user pages, (2) supporting the Korean
Figure 5. Enlarged view of the Terri Shiavo page [43] in Revert claims in discussions on the users’ talk pages, (3) preferring the
Graph. (a) User first clicks on a specific user node; (b) Revert term “East Sea” over “Sea of Japan”, (4) preferring “Dokdo” over
Graph shows a list of revert relationships that Alienus is involved in; the alternate “Takeshima” or the more neutral “Liancourt Rocks”.
(c) Clicking on a specific relationship in the list shows only the We observed users in Group C disputing the points of view of
revert history between those two users. Group A. This group includes users who (1) dispute the official
U.S. position (which supports the Korean occupation), (2) openly

167
 refute the Korean point-of-view on their talk pages, as well as 3) 6.4 PATTERN THREE – FIGHTING VANDALISM
directly claim Japanese heritage or affiliation on their user pages. Fighting vandalism is an issue in Wikipedia. As shown in Table 1,
Unlike Group A and C, users in Group B showed mixed about 24% of total reverts are made due to vandalism. Revert
opinion on the issue. Group D, depicted only in the figure but not Graph uncovers clear patterns of vandalism and anti-vandalism
tallied in the table, is not considered in this analysis because they efforts. As shown in Figure 7, two user groups emerge in each of
primarily have very short edit histories and 31 out of these 34
these examples, where users in one group reverts revisions from
users are non-registered users.
the other group. This first group usually includes many
Our analysis is summarized in Table 2, which shows that the
administrators (green nodes) while the second group often
identified user groups indeed represent distinct opinion groups.
contains mostly anonymous users (white nodes).
Number of users in user group A B C Total
Users with Korean point of view 10 6 0 16
Users with Japanese point of view 1 8 7 16
Neutral or Unidentified 7 3 6 17
Table 2. User Groups on the Dokdo article.

As shown in this conflict pattern, node clusters identified in

Revert Graph correspond to opinion groups that are not only
cohesive but also represent major points of view in these topic
areas.

6.3 PATTERN TWO – MEDIATION

Another common pattern revealed by Revert Graph is a group of
users attempting to mediate between user groups with divergent
points of view. These users are not active in expressing a Figure 7. Examples of the vandalism pattern in Revert Graph. The
particular view, but rather trying to moderate the discussion left is Revert Graph applied on Acid rain page [33] and the right is
towards a neutral version of the article. Users in these mediator on Uranus page [45]. Two types of user groups are clearly
roles usually revert edits from many other user groups. This group identifiable in each graph. The majority of the upper groups are
rejects revisions that do not satisfy Wikipedia content policies anonymous users and vandalized articles. The bottom group, which
such as neutral point of view [41] and attribution [35]. For mostly consists of administrators, are fighting vandalism.
example, in the Terri Shiavo controversy depicted in Figure 6, In addition, we found that an anti-vandalism robot Tawkerbot2
users in the center group, which is marked by a dotted circle, [34] is often actively engaged in this pattern. The left image of
revert edits made by other user groups. Their comments revealed Figure 7 shows that the robot (indicted by a dotted circle) does a
that they believe those edits did not satisfy Wikipedia policies. decent job of fighting vandalism.

6.5 PATTERN FOUR – CONTROVERSIAL EDITOR

In Revert Graph, the size of a user node can be mapped to the
number of reverts or revisions. In addition, the thickness of the
edges represent the degree of revert relationships between users.
Together, these features increase the visual saliency of editors
who engaged in many reverts, providing an intuitive way to
quickly identify users who may be involved in high degrees of
conflict. For example, Figure 8 shows how certain controversial
editors are easily identifiable in the Revert Graph. They are
usually self-appointed experts, or have strong points of view. In
the left side example, a few editors appointed themselves as the
ultimate source of knowledge relating to Darth Vader. In the right
side example, several editors expressed strong personal opinions
on Windows Vista.

Figure 6. A mediator pattern in Revert Graph applied on Terri

Shiavo page [43]. Users in the center group (marked by a dotted
circled) revert edits by users in other user groups.

User nodes in Revert Graph are color-coded based on their

roles, such as whether they are administrators. Therefore, it is
easy to identify that mediators are often administrators as in
Figure 6. Figure 8. Controversial editors are visually identifiable in both
Revert Graphs. Left: Darth Vader page [38]. Right: Windows Vista
page [46]

168
 As another example, editors in Figure 1 are involved in debates view to an article can be controversial because it can be hard to
on Charles Darwin. These editors with heavy revert relationships determine whether it improves the neutral point of view of the
are visually salient and easily identifiable. article or if it is just vandalism. Sometimes we found it hard to
determine this difference ourselves. Improved tools for situation
7 DISCUSSION awareness, such as Revert Graph, should enable editors to quickly
The Revert Graph visualization helps identify important social get a gist of these kinds of conflict patterns.
patterns in Wikipedia such as groups with differing viewpoints, 8 LIMITATIONS
mediation, vandal fighting, and user conflicts. The above
scenarios demonstrate the effectiveness of the revert-based As shown in the case studies, there are various types of
conflict model, which is based on simple revert relationships disagreements and conflicts in Wikipedia. One limitation of our
between users. These results show how surfacing information tool is that, while some conflict patterns are discovered (i.e.,
about user behaviour built up through natural system usage can vandalism, mediation, etc), not every aspect of social dynamics in
lead to valuable insights that can help make sense of the evolution online collaboration systems was fully addressed. For example,
of pages and the motives of users. sources of disagreements, types of conflicts, and motivation for
We believe this paper may help shape the development of editing are deeper questions that require further research. The
conflict resolution tools, which are expected to play an important answer for those questions would provide useful guidelines for
role in online collaborative systems. These results are most designers of online collaborative communities.
directly applicable to other wiki-based systems in which reverts Another limitation is that the force-directed layout does not
are tracked as part of system usage. Wiki systems are being always produce optimal user groups, because it requires sufficient
deployed in a wide variety of domains, including intelligence revert relationships in the data set. Since Revert Graph relies
analysis [27], corporate memory [28], and scholarly research [26]. solely on revert relationships, the tool cannot detect conflicts
Tools such as Revert Graph could be useful in quickly identifying between users who were not involved in reverts.
differing opinion groups and promoting situational awareness.
9 CONCLUSION
The user conflict model and Revert Graph techniques described
here may be applicable to other online communities where Visual analytics can provide useful tools for users to make sense
conflict relationships between users can be identified. We of the state of complex collaborative environments. Wikipedia
demonstrated that it may be possible to identify conflicts and and wiki-style spaces serve as important examples of how
viewpoints by using negatively-valenced relationship data complex collaborative environments develop and evolve over
collected automatically through normal usage of a system. We time. The rapid growth of Wikipedia presents a challenge for
believe that the principles behind Revert Graph may be applicable analysts to understand conflicts and other social dynamics.
to other, non-wiki based systems where identifying conflict, user To address this challenge, we have been building a model of
factions, and shared viewpoints is important. Some examples how conflicts occur in Wikipedia and how conflicts are resolved.
include detecting and visualizing collusion and user factions in We used this model to develop a visualization tool called Revert
social collaborative systems where content or users are rated (e.g., Graph to facilitate analysts to understand user conflicts and
digg.com, Slashdot.org). explore opinion groups revealed in the visualization. We also
In addition to the four social patterns described above, we also demonstrate the effectiveness of the tool by presenting interesting
offer some interesting observations that might be useful in user patterns and case studies that we found in Wikipedia.
designing these tools. First, we found that there were unexpected The tool can answer questions such as the severity and form of
controversial sub-topics in Wikipedia articles. For example, in the the disagreement as well as the shape and size of opinion groups.
Charles Darwin page [37], we observed that several users were in The case study revealed conflict patterns such as the identification
dispute over a trivial detail relating to the fact that Abraham of (a) the formation of opinion groups, (b) patterns of mediation,
Lincoln and Darwin share the same birthday. To our surprise, (c) vandalism, and (d) major controversial users and topics.
“September 2005” was a controversial article. We found that high Further research is needed to understand how visual analytics
degrees of conflict in that article were caused by news articles can help analysts to understand a wide variety of problems in
reporting disputes in Middle East countries. Discovering these collaborative spaces. We believe that the approach taken here
types of unexpected conflicts would be a hard task without the aid may inform the development of other conflict resolution and
of Revert Graph. situation awareness tools for collaborative environments.
Second, disagreements between groups of users often spill over
to other related articles. For example, during our investigation of ACKNOWLEDGEMENTS
the Dokdo article, users who made reverts on the Dokdo page
were also found to have made reverts on other Korean-Japanese We would like to thank Peter Pirolli, Stuart Card, and Mark Stefik
pages such as the “Sea of Japan” page. Some users involved in for valuable advice, and to James Forrester who helped facilitate
conflicts in the “Anarchism” page also joined the disputes in the this research.
“Anarcho-capitalism” page. This observation strongly implies the
usefulness of a topic-based conflict model. REFERENCES
Third, we observed various social processes for reaching [1] Adamic, L.A., Glance, N. (2005) The political blogosphere and the
consensus. As well as using article talk pages and user talk pages, 2004 U.S. election: divided they blog, Conference on Knowledge
Wikipedians also 1) discuss with third parties, 2) conduct surveys, Discovery in Data, Proc. 3rd International workshop on Link
3) request mediation, and 4) call for arbitration [42]. Among discovery, 36-43
them, we found that user talk pages were very useful in [2] Ahrens, F., A Wikipedia for Secrets, Washington Post, November
understanding disagreements and conflicts between users. We 5th 2006.
believe further effort in modeling user talk pages would provide a [3] Arguello, J., Butler, B. S., Joyce, L., Kraut, R., Ling, K. S., Rosé, C.
deeper insight into the social dynamics of Wikipedia. P., et al. (2006). Talk to me: Foundations for successful individual-
Fourth, editors often get confused between “neutral point of group interactions in online communities. In CHI 2006: Proceedings
view” and “vandalism”. For example, introducing a new point of

169
 of the ACM conference on human-factors in computing systems (pp. [22] Noel, S., and Robert, J. M. How the Web is used to support
959 - 968). New York: ACM Press. collaborative writing. Behaviour & Information Technology, 22
[4] Brands, U., Fleischer, D., Lerner, J. (2006) Summarizing Dynamic (2003) 245-262.
Bipolar Conflict Structures, IEEE Trans. On Visualization and [23] Perer, A. and Shneiderman, B. Balancing Systematic and Flexible
Computer Graphics, 12(6), 1486-1499. Exploration of Social Networks, IEEE Symposium on Information
[5] Buriol, L., Castillo, C., Donato, D., Leonardi, S., and Millozzi, S.: Visualization (Infovis 2006)
"Temporal Evolution of the Wikigraph", Proc. Web Intelligence [24] Pirolli, P. Information Foraging Theory: Adaptive Interaction with
Conference 2006. IEEE CS Press (2006), 45-51. Information. 2007. New York, NY: Oxford University Press.
[6] Capocci, A. Servedio, V.D.P., Colaiori, F., Buriol, L.S., Donato, D. [25] Rivest, R., The MD5 Message-Digest Algorithm, RFC 1321, MIT
Leonardi, S., and Caldarelli, G., Preferential attachment in the and RSA Data Security, Inc., April 1992.
growth of social networks: the case of Wikipedia, [26] Scholarpedia, http://www.scholarpedia.org
arXiv.org:physics/0602026 (2006). [27] Shrader, K., Over 3,600 intelliegence professionals tapping into
[7] Cosley, D., Frankowski, D., Terveen, L., Riedl, J. Using Intelligent “intellepedia”, USA Today, November 2nd 2006.
Task Routing and Contribution Review to Help Communities Build [28] Socialtext, http://www.socialtext.com
Artifacts of Lasting Value. Proc. CHI 2006, ACM Press (2006), 22- [29] Thomas, J. J. and Cook, K. A. (2005) Illuminating the Path: The
27. Research and Development Agenda for Visual Analytics. IEEE CS
[8] Curtis, P. Mudding: Social phenomena in text-based virtual realities. Press.
Palo Alto Research Center, 1992. [30] Viégas, F. B., Donath, J. S. (1999) Chat Circles, Proc. CHI 99, ACM
[9] Dibbell, Julian. A Rape in Cyberspace. Village Voice, 21 (1993), 26- Press, 9-16.
32. [31] Viégas, F. B., Wattenberg M., Dave K., Studying cooperation and
[10] Dourish, P., & Bellotti, V. Awareness and coordination in shared conflict between authors with history flow visualizations, Proc. CHI
workspaces. Proc. ACM Conference on Computer Supported 2004, ACM Press (2004), 575-582.
Cooperative Work (CSCW ‘92), ACM Press (1992), 107-114. [32] Voss, J. Measuring Wikipedia. Proc. 10th Intl. Conf. Intl. Soc. for
[11] Endsley, M. (2003). Designing for Situation Awareness. Boca Scientometrics and Informetrics (2005).
Raton, FL: CRC.Press. [33] Wikipedia.org, Acid ran, http://en.wikipedia.org/wiki/Acid_rain
[12] Flor, N. V., Hutchins, E. (1991): Analyzing distributed cognition in [34] Wikipedia.org, AntiVandalBot,
software teams: A case study of team programming during perfective http://en.wikipedia.org/wiki/User:Tawkerbot2
software maintenance. In: Proc. of the 4th Annual Workshop on [35] Wikipedia.org, Attribution, http://en.wikipedia.org/wiki/WP:A
Empirical Studies of Programmers. 1991, Norwood, New Jersey, [36] Wikipedia.org, Be bold in updating pages,
USA. p.36-59. http://en.wikipedia.org/wiki/WP:BB
[13] Franco, V., Piirto, R., Hu, H. Y., Lewenstein, B. V., Underwood, R., [37] Wikipedia.org, Charles Darwin.
and Vidal, N. K. Anatomy of a flame: conflict and community http://en.wikipedia.org/wiki/Charles_Darwin
building on the Internet. Tech. and Society Magazine, IEEE, 14 [38] Wikipedia.org, Darth Vader,
(1995) 12-21. http://en.wikipedia.org/wiki/Darth_Vader
[14] Giles, G. Internet encyclopaedias go head to head. Nature, 438, 7070 [39] Wikipedia.org, Dokdo. http://en.wikipedia.org/wiki/Dokdo
(2005), 900-901. [40] Wikipedia.org, Edit war,
[15] Hadoop Project. http://lucene.apache.org/hadoop/. http://en.wikipedia.org/wiki/Wikipedia:Edit_war
[16] Heer, J., Card, S. K. and Landlay, J. Prefuse: A toolkit for [41] Wikipedia.org, Neutral point of view,
interactive information visualization, Proc. CHI 2005, ACM Press http://en.wikipedia.org/wiki/WP:NPOV
(2005), 421-430. [42] Wikipedia.org, Resolving Disputes,
[17] Heer, J., Boyd, D., "Vizster: Visualizing Online Social Networks," http://en.wikipedia.org/wiki/Wikipedia:Resolving_disputes.
infovis, p. 5, Proceedings of the 2005 IEEE Symposium on [43] Wikipedia.org, Terri Schiavo.
Information Visualization (INFOVIS'05), (2005) http://en.wikipedia.org/wiki/Terri_Schiavo
[18] Kittur, A., Suh, B., Chi, E. H., Pendleton, B. A. He says, she says: [44] Wikipedia.org, Three-revert-rule.
Conflict and coordination in Wikipedia. CHI 2007; 2007 April 28 - http://en.wikipedia.org/wiki/Wikipedia:Three-revert_rule
May 3; San Jose; CA. [45] Wikipedia.org, Uranus, http://en.wikipedia.org/wiki/Uranus
[19] McGrath, J.E., 1991. Time, interaction, and performance (TIP). A [46] Wikipedia.org, Windows Vista,
theory of groups. Small Group Research 22 (2), 147-174. http://en.wikipedia.org/wiki/Windows_Vista
[20] Nardi, B. A. and Miller, J. R. 1991. Twinkling lights and nested [47] Zlatic, V., Bozicevic, M., Stefancic, H., Domazet, M. Wikipedias:
loops: distributed problem solving and spreadsheet development. Int. Collaborative web-based encylopedias as complex networks.
J. Man-Mach. Stud. 34, 2 (Feb. 1991), 161-184. Physical Review E 74 (2006).
[21] Neuwirth, C. M., Kaufer, D. S., Chandhok, R., and Morris, J. H.
Issues in the design of computer support for co-authoring and
commenting. Proc. CSCW 1990, ACM Press (1990), 183-195.

170