Danyel FisherMicrosoft Research

Using Egocentric Networks to Understand Communication

Social-network analysis generally helps researchers understand how groups of

people interact. The author uses small-scale egocentric social networks, based

on volitional, explicit connections, to understand how people manage their

personal and group communications. Two research projects using this approach

show that such networks can give researchers important insight into the people

who communicate online. Soylent, a project based on email, shows several

common patterns in social interaction. The Roles project, based on Usenet

newsgroups, suggests that various online social spaces can behave very differently

from each other.

We can learn a lot about peoplefrom who they talk to — whetherthey talk to sociable, well-

connected people or to unconnected indi-viduals, for example. Is the person part ofa large group whose members all talk toeach other, or does he or she bridge socialworlds? Examining such connectionsbetween people can teach us how theyoperate socially.

Online communication, in particular,is an increasingly important way to getinformation from, and keep in contactwith, each other. Communication througha computer can be recorded and thenanalyzed; we can later use that analysisto develop systems that are more awareof how people interact online.

Although much social-network analy-

sis attempts to examine complete, large-scale networks — ones in which everyconnection between network members isvisible — my work takes a more restrictiveview. Using egocentric network-analysistechniques — which examine only peo-ple’s immediate neighbors and associatedinterconnections — helps us learn abouthow individuals correspond with theirsocial networks.

I’ve used this approach in two differ-ent research projects: the Soylent project,which I conducted as a graduate studentat the University of California, Irvine,explores and describes interaction pat-terns in email.1 The Roles project, con-ducted at Microsoft Research, appliessocial networks to Usenet messages.2

In both cases, egocentric networks of

20 SEPTEMBER • OCTOBER 2005 Published by the IEEE Computer Society 1089-7801/05/$20.00 © 2005 IEEE IEEE INTERNET COMPUTING

Soci

al N

etw

orki

ng

immediate connections provide network analystsand researchers with interesting and useful dataabout the roles individuals adopt and the typesof interactions that develop among their socialpartners.

A Perspective on Social NetworksAs in all network analyses, the networks in thisstudy comprise interconnected nodes and edges— an edge connected to a pair of nodes representsa pair of people and the relationship that linksthem. We can measure these networks quantita-tively, using techniques from both sociology andgraph theory to describe people’s interactions asrecorded within the network. Network analysisthen gives us a structural description of thegroup’s relationships.

The networks in both the Soylent and Rolesstudies are somewhat different from many networkanalyses studied in other contexts. First, the edgesin these networks are volitional — that is, everyedge is motivated by a specific social choice. Suchchoices are linked to the people with whom indi-viduals interact, and network ties are positive evi-dence of meaningful relationships. Although socialconstraints and local norms might drive these rela-tionships, they are intentional. Conversely, thenonvolitional networks that many network analy-ses study might measure more unintentionaleffects — examining edges such as “has sat next toon an airplane,” for instance — and can collectinformation about mobility and proximity (whichmakes them very useful for studying disease trans-mission, media exposure, and so on). However,they have a limited ability to measure social choiceand preference.

The particular edges that Soylent and Rolesexamine are based on personal communication.Each edge is based on at least one message, sentfrom a single person and directed to another.Additionally, each message represents time andeffort — that is, a decision that the message isworth sending and that it’s worth sending to aspecific person.

Unlike with many network analyses, I don’tassume that networks in the Soylent and Rolesprojects are transitive in any meaningful way.Many studies identify “gatekeepers”3 who controlflows such that information can move only fromA to C via B. For example, a salesman (B) mightcarefully guard his contacts (C) from his col-leagues (A). In the systems I discuss, this doesn’tnecessarily occur. That A connects with B and B

connects with C doesn’t mean that B is a middle-man between the two. Indeed, in these networks,A could talk directly with C as desired: the lackof a direct connection implies that A chooses notto contact C. In the case of Usenet newsgroups,for example, even the most casual reader canaccess all posters’ names; anyone can reply to aprevious message, thus forging the ties that thestudy observes.

Seeing from One ViewpointIn addition to these constraints, the Soylent andRoles projects share two other important themes.First, they use only information that’s available tothe system and to the user. Many social-networkanalyses present a view of networks based on datathat no single person could know. Frequently,they’re based on extensive data collection or priv-ileged access to servers: they list all the connec-tions within a single organization or computernetwork. Such projects can discover previouslyunknown information that no single person canfind out. In one research project from Hewlett-Packard laboratories, for example, the authorsidentified some of the informal communities with-in their organization.4

Soylent and Roles aggregate information that’snot necessarily novel — users can already accessthe information, although they can’t ordinarilyview it from a network-based perspective. Aggre-gating this known data is useful initially for devel-oping information-visualization designs that canpresent users with information in new ways (seethe “Network Visualizations” sidebar for moreinformation). The insights we can draw from thesenetwork analyses can also lead to the developmentof new user-interface tools.

Additionally, Viegas and colleagues argue thatshowing personal information to users lets themnavigate their social histories, examining whothey’ve encountered and what experiences they’vehad. Sharing a social network, therefore, can belike sharing a photo album.5 Indeed, both Viegasand Nardi and colleagues6 note that, for manyusers, visualizations of personal networks becomea locus of storytelling.

Finally, using only information that is avail-able to users has beneficial privacy trade-offs. Ina traditional whole-network study, researchershold valuable and potentially dangerous informa-tion that an individual might be afraid to exposeto researchers or other participants. When usingonly local data, we lose a global scope, but need

IEEE INTERNET COMPUTING www.computer.org/internet/ SEPTEMBER • OCTOBER 2005 21

Egocentric Networks

22 SEPTEMBER • OCTOBER 2005 www.computer.org/internet/ IEEE INTERNET COMPUTING

Social Networking

not worry about the potential harm of exposingprivate data. Soylent presents personal email tousers via a tool that lets them review only theirown data; no one else needs access to it. The Rolesproject examines groups that are publicly archivedby Google. By presenting only information thatusers can control, we make it more comfortable forthem to participate in the study, and make the datapotentially less harmful — they need not fearexposing private information.

The second important aspect of my approach isthat limiting the examination to a single-personperspective generates several salient questions forresearchers — who is this person connected to, forexample, and how are his or her neighbors inter-connected? These, in turn, have design implica-tions: a person connected to a sparse networkmight be particularly interested in a tool that helpsjuggle different contexts; a person linked into adense network might prefer a tool that puts theirinformation in one place.

In the Soylent project, for example, we startwith the system’s user (called Ego), who has sentemail to a number of different people. AlthoughEgo presumably knows something about whothese people are, his email client doesn’t necessar-ily make this context accessible. Even if it storesthe fact that Ego has communicated with them in

the past, it’s unlikely to have stored informationwith any depth, such as who initiated the last con-versation, when it happened, or who else wasinvolved. We focus, then, on Ego’s neighbors, oneat a time, trying to understand their connectionsto each other and to the entire network. In theRoles project, we focus on one member of eachnewsgroup at a time, trying to understand theirparticular roles.

The Soylent ProjectIn today’s computer systems, “people” appear onlyincidentally, as disconnected names in addressbooks, contacts in mailing lists, or documentauthors. They exist out of context with one anoth-er. The Soylent project, which I conducted from2002 to 2004 as part of my dissertation work at UCIrvine, aims to rethink the way we currently useboth email and desktop space. My goal with theproject was to connect people to each other and totheir documents — placing people and projectswithin a shared social context. I believe this is avaluable direction for future user interfaces. Today,most systems separate documents from theirauthors: documents are stored in file systems,whereas information about people is stored inemail and contact-management tools, requiringusers to manage the two separately. By storingboth in a shared space, it’s possible to manage andsimplify this complexity.

Soylent analyzes personal social networksfrom email archives, looking only at outgoingemail from a single user at a time. Basing the net-work entirely on outgoing mail lets us view users’perceptions of their social worlds. Althoughincoming mail might be spam, mass emails, orirrelevant messages, we can assume that eachmessage that users send is relevant to them — atleast when they send it. Based on a user’s activi-ty history, Soylent ultimately assembles a net-work of connected people that closely resemblesthat of Social Network Fragments5 and PersonalMap,7 both of which plot the interconnectionsbetween a user’s contacts and present this infor-mation to the user.

By examining the network developed by mes-sages co-addressed to two or more contacts, wecan determine what social networks the user con-siders relevant. Many email messages from me toboth Gina and Phil, for example, suggests that Ibelieve Gina and Phil have common interests. ThatI infrequently (if ever) send email addressed toboth my brother and my officemate might suggest

Figure 1. A schematic view of ego-centric networks, based onoutgoing mail. Soylent assembles three email messages from A into anetwork that shows the connections between A’s correspondents.

From: A@here.nowTo: X@uci.edu, Y@berkeley.eduDate: Monday, 4:00 pm

From: A@here.nowTo: Y@berkeley.edu, Z@uci.eduDate: Wednesday, 12:00 pm From: A@here.now

To: W@boeing.com, V@ibm.com, Z@uci.eduDate: Monday, 11:00 am

X Y

W

V

Z

A

that they inhabit very different social worlds. Thisnetwork — the connections between my alters (orneighbors) — represents an ego-centered perspec-tive on how email contacts are interrelated.

I ran the tool on 15 users’ machines, generat-ing and exploring their networks, and then askedthe users to describe what sorts of connectionsand networks they observed. Who was connect-ed, for example, and where did those connectionscome from? I observed commonalities in the net-works that came out of these users’ mail. Theusers explained the networks, telling stories aboutthe connections — they identified the cluster rep-

resenting their soccer team, for example, or theties between marketers and designers at theircompanies. Figure 2 shows the four core patternsthat I interpreted as structural patterns from cer-tain recurrent roles that emerged from the pro-ject. We can view these images from Ego’sperspective — Ego does not appear in the images,but each pattern shows how his neighbors areconnected to him.

The disconnected clusters pattern occurs whengroups of people aren’t connected to each other atall. In Figure 2a, small social clusters are visiblearound the outside of the display, whereas a large

IEEE INTERNET COMPUTING www.computer.org/internet/ SEPTEMBER • OCTOBER 2005 23

Egocentric Networks

Figure 2. Four network patterns. (a) Disconnected clusters occur when groups of people aren’t at all connected. (b) The onionpattern shows a tightly connected group at the center of a loosely connected group. (c) In the nexus pattern, a single personshares multiple contexts with Ego. (d) The butterfly pattern occurs when two different groups are connected by one person.

Okwms Zafryql

xuflallzBplxa Zaawa

xuqk@oja.enu.ssk

Vnump Jpasdfrbhue

nukakal

íxpkdm@cls.qqií

tvhjatpm.bzqqsg

Eaymi Ymmwtxrg

Eayi Mvopsy

dufa7670-nzogcgac

Naglz Ymmnj

Nauo Hlzpwaz

fmaxjxpx

Djavquabn Wqrajpai

Tvhzjt E. Egghau

tfe

jthrj.zqpsdybbsy

Kafo Sbzawaffi

nawxnaz.ncmvbfcijgg

Dnhmiyvv Bzzazs Sgmygdl Ce Uryea djumx.oqwnd

Tjuidaz Mydtv

tyw

ltlwHnrvna Uabv

uaoiw, rrvwXaqha Wzade

Sqeiwzv Bswha Uyczh

tm v eaymi.wqwowxr

Ftlr Hyamxpd

Xabrj Sccdpdf

adksw

kqqaynsu

atymsjr

Sqvm

Tvhzj Naug Ytqcr

YitwEaqam

Gywhyqmqp

Goasol

Ctbas

(c) (d)

(a) (b)

central cluster shows a tighter set of interactions.Ego connects many disconnected groups that havenothing to do with each other. Although thisimage suffers from the crowding discussed in thesidebar, it’s meant to illustrate only one feature:the clusters’ disconnected nature.

In the onion pattern, a tightly connected groupemerges at the core of a loosely connected group.Ego is a participant in the core group. This patternwas common in many circumstances, includingthose in which we found a small project team atthe core of a larger set of interested parties — con-tractors, administrators, and so on. In Figure 2b,for example, we see the presenters at a conferencesurrounded by the larger network of attendees.

In the nexus pattern, a single person sharesmultiple contexts with Ego. In Figure 2c, we see acoworker who cooperated with Ego on variousprojects. The nexus was also common amongmanagers and administrators — people repeatedlyin touch with Ego even as Ego shifted among dif-ferent contexts.

The butterfly approach occurs when two differ-ent groups of Ego’s neighbors overlap via a singleperson. In Figure 2d, for instance, the one personwas involved in Ego’s research group and also ina workshop that Ego arranged.

These network patterns suggest that differentcontexts differ visibly as well. The patterns aredetectable, recognizable, and indicative of the dif-ferent sorts of online relationships that peopleshare. In other words, the social contexts in whichpeople are embedded are, in fact, ones that areworth recognizing.

But what can we do with these patterns, oncewe’ve identified them? We might be tempted tostraightforwardly cluster the nodes from the net-work, breaking them into lists based on which peo-ple are most connected.8 Each person might beplaced in one list (such as “coworkers” or “friends”)based on their nearest neighbors. Yet, the patterns’complexity argues against such a simplisticapproach. The onion pattern, for example, demon-strates that both inner and outer circles of partic-ipation often exist; thus, any possible solutionmust be able to adjust its lists to different levels ofintimacy. The butterfly pattern shows that we can’tsimply characterize a person into a single contextat a time; rather, our relationships are multifacetedand might be tied to various concerns. The nexuspattern similarly suggests that even strong collab-orators operate in a wide range of contexts. Anemail from a collaborator might be associated with

one of many different projects; knowing some-thing about the range of projects that the collab-orator is involved in might help an individual sortincoming email into the correct category.

Interfaces for handling contacts and commu-nication should be lightweight and supportive ofdynamic groups. We can more easily build inter-faces that support a notion of “relevant people” ifthey maintain that notion dynamically, based onsocial structure. We might build systems that focuson not just the person to whom a particular emailis addressed but also the several sets of people towhom the recipient is most closely associated.

Roles in ConversationI conducted the Roles project with MicrosoftResearch’s Community Technologies Group in2005, looking at archived data from 2001 and2004. Microsoft Research has had a continuinginterest in social behavior in online spaces.2 Onepersistent question has involved identifying thegood answers in question-and-answer orientednewsgroups. To do this, we must know somethingabout who asks and who answers questions with-in these newsgroups. The Roles project looks at thesocial roles that emerge among Usenet newsgroupmembers and uses structural data around thesegroups to help us understand how contributors toonline social spaces differ from each other.9

My own contribution was to bring a social net-work approach to the Usenet data. I based thisresearch on a replied-to network that specifieswhich authors have replied to which others withina given newsgroup. I then examined these net-works to see whether authors preferred to reply tosome people rather than others, and whether dif-ferent sorts of authors or different newsgroupsexhibited characteristic patterns.

Usenet is an interesting testbed for this pro-ject. A 25-year-old online distributed discussionand posting space, it predates blogs, Web-baseddiscussion boards, and even the Web itself.Although less prominent today than it once was,Usenet continues to have substantial traffic:hundreds of millions of messages are posted,read, and sorted into tens of thousands of topi-cal newsgroups annually. One particularly inter-esting aspect is that all groups receive the sameinterface. Whether a newsgroup dedicated totechnical questions and answers or one devotedto religious discussion, users have a similar expe-rience — they can read messages in threads, postnew messages in response, or start new threads.

24 SEPTEMBER • OCTOBER 2005 www.computer.org/internet/ IEEE INTERNET COMPUTING

Social Networking

IEEE INTERNET COMPUTING www.computer.org/internet/ SEPTEMBER • OCTOBER 2005 25

Egocentric Networks

This commonality allows for a broad compara-tive analysis.

Within this uniform space, different behaviorand roles emerge. A politically oriented newsgroupwill exhibit very different behavior than a social-support newsgroup, for example. The Roles projectattempts to distinguish some of these roles usingsocial-network tools.

For given newsgroups over a given month, Iextracted social-network information from Micro-soft’s Netscan archive10 by collecting all occasionsin which one person replied to another. I then cre-ated an ego-centric network around participants inthese newsgroups. For each of these hundreds ofnetworks per newsgroup, a single node represent-ed an individual author, whereas directed edgesrepresented the number of times that one authorreplied to another. Thus, an edge from A to B withweight 3 indicates that A replied to a message fromB three times. The network was restricted to twodegrees out from the center: we saw the participant,the people who the participant interacted with, andthe people who they, in turn, were connected to.

My research group extracted networks repre-senting a month’s worth of replies from nine dif-ferent newsgroups on five different topics. (Wesampled eight of the nine groups through Janu-ary 2001, and microsoft.public.windows.server.general during November 2004). Table1 lists the newsgroups we sampled and their gen-eral topics. To demonstrate how large the news-groups were, it also lists the number of replies(that is, messages involved in conversation) anddistinct authors during that time period.

We wanted to identify “answer people” — thosewho are instrumental in answering technical ques-tions posed in technical newsgroups. Although fewnewsgroups overall are technical, such groups areparticularly interesting for understanding the com-munity that grows around software, and forobserving a highly active user population. Answerpeople are rare, but distinctive, and we can recog-nize them using local social-network data. To doso, we examine both an individual’s out-degree —the number of different people to whom they’veresponded — and those of their immediate neigh-bors. Answer people tend to respond to those whodon’t respond to many others — that is, they havea high out-degree, but are largely connected tothose with low out-degrees.

We can summarize this behavior with out-degree histograms, which show the distribution ofthe degrees for any given person’s neighbors. This

measure lets us infer how a person interacts withothers. For example, a heavily left-weighted his-togram is common to people offering technicalsupport. Figure 3a (next page) shows an answerperson from the group microsoft.public.windows.server.general. The answer person, atcenter, replies to several people who are themselveslargely disconnected from anyone else — that is,those who are just starting conversations or whohave responded to only a small group of people. Inthis example, the answer person also replies to oneperson with a high out-degree, illustrating thatanswer people occasionally clarify answers foreach other or disagree with another’s answers.

In contrast, a more heavily right-weighted his-togram occurs with people who carry on prolongedand frequent discussions, such as members of apolitical newsgroup. Figure 3b shows a networkdiagram and a histogram for the group alt.politics, which is based on brisk discussion ofcurrent political issues. Although this stripped-down network diagram is crowded, we can seeseveral important differences between it and thetechnical support group’s diagram. Most obvious-ly, the central core is tightly interconnected. Unlikein the technical support group, tightly connectedpeople reply to other connected people.

Figure 3c shows a characteristic member of oneflame group we observed. Flaming involves vocal,often insulting, conversations within groups.11 Thisdiagram looks somewhat similar to the politicaldiscussion group’s, with one exception — conver-sation in flame groups has a very strong recipro-cal aspect. Whereas we traditionally see reciprocityas a valuable aspect of conversation, in onlinegroups, it frequently indicates flame wars. Our datashows flame wars to be concentrated around pairsof people — although many people might partici-pate at the beginning, most of the conversation

Table 1. Newsgroups sampled during the Roles study.

Topic Name Replies AuthorsTechnical support comp.soft-sys.matlab 712 437

microsoft.public.windows.server.general 1,489 855Discussion rec.kites 924 276

rec.music.phish 6,105 1,263Political discussion alt.politics 16,059 2,187Flame alt.flame 3,802 618

alt.alien-vampire.flonk.flonk.flonk 6,494 755Social support alt.support.divorce 2,937 339

alt.support.autism 3,095 188

26 SEPTEMBER • OCTOBER 2005 www.computer.org/internet/ IEEE INTERNET COMPUTING

Social Networking

Figure 3. Typical roles in four different newsgroups. We examined groups for (a) technical support, (b)political discussion, (c) flaming, and (d) social support. The left side shows a given contributor’simmediate neighbor networks; the right side shows the out-degree distributions for each group. Out-degree distributions are on both a logarithmic x-axis, so that each bucket is twice as large as theprevious one, and a y-axis, so that a column one unit higher represents roughly twice as many neighbors.

6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0

(d)

(c)

(b)

(a)

6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0

6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0

6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0

Num

ber

of n

eigh

bors

Num

ber

of n

eigh

bors

Num

ber

of n

eigh

bors

Num

ber

of n

eigh

bors

0 1 2–3 4–7 8–15 16–31 32–63 64–127 128–256

Range of degrees

0 1 2–3 4–7 8–15 16–31 32–63 64–127 128–256

Range of degrees

0 1 2–3 4–7 8–15 16–31 32–63 64–127 128–256

Range of degrees

0 1 2–3 4–7 8–15 16–31 32–63 64–127 128–256

Range of degrees

IEEE INTERNET COMPUTING www.computer.org/internet/ SEPTEMBER • OCTOBER 2005 27

Egocentric Networks

occurs between two discussants by the end. Indeed,the figure shows several thick lines around the coremember, indicating his flame-war history.

Finally, a substantially different characteristicpattern exists in groups dedicated to social sup-port. Figure 3d shows the typical out-degree dis-tribution within alt.support.divorce — themost active people have a bump at both the lowand high ends. This suggests that they bothrespond to the requests of new participants look-ing for support and continue to talk with the mostactive members.

Interestingly, in none of these groups can weexplain the distribution merely using the partici-pants’ raw in- or out-degrees. That is, their activ-ity levels don’t correlate with their roles. All of thegroups share the common power-law-like distrib-ution of in- and out-degrees: a large number ofalters to whom very few messages were sent, witha long tail of few alters to whom far more mes-sages were sent.

Although our work is still in progress, it leadsus in some intriguing directions. Distinct behav-iors emerge in different newsgroups: alt.flame hasno answer people, for example, whereas alt.politics appears to have no social-support peo-ple. On one level, this is unsurprising — the groupshave, after all, different goals and needs. However,it’s extremely surprising when we consider howfew formal constraints exist to regulate how actorsbehave in these groups (posters can’t be kicked offor fined for misbehavior, for example). Althoughall the newsgroups’ interfaces look the same, rad-ically different roles emerge in the spaces.

The perspectives of both the Soylent and Rolesprojects are intentionally limited. In many

cases, examining a whole group or full networkcan lead to interesting and useful results. Howev-er, each of those approaches runs into issues ofboth complexity and privacy. Our combination ofbuilding on public data and maintaining an imme-diate perspective seems like a natural and valuableway to study online communication.

Given that the Solvent project studies only out-going email, it fails to include other resourcesthrough which people collaborate, including doc-uments, presentations, other email accounts, andface-to-face contact. To broaden this project’sscope, we could annotate all documents within acomputer system with social data, such as where(and whom) it came from and who has edited it,

sent it, or received it. Then we could broaden thepatterns I’ve just discussed to include artifacts aswell as people.

In the Roles project, we might ultimately beable to automatically detect and classify the dis-tinct behaviors that emerge among users. Userscould thus discover whether the person they’rereading tends to take the role of an authority ora flamer. An enhanced search engine could pro-vide contextual information — are the resultsfrom someone who provides social support ordiscusses issues? Different audiences might findthis information valuable in different contexts.The techniques just suggested — using social net-work data — could expand the ways in which weuse social metadata to understand and reflectgroup behavior.

It would also be valuable to track the stabilityof both groups and the roles played by individu-als. Several questions emerge once we think of anewsgroup as inhabited (and, in part, defined) bya set of roles within it. How rapidly do individualswithin a single group change? Are their roles con-sistent between different newsgroups (suggestingthat people search for newsgroups that match theirstrengths), or do they adapt to the environmentsin which they find themselves? How firmly donewsgroups hold on to these roles, and under whatcircumstances do they change? Our work atMicrosoft continues to develop and explore thesequestions further.

AcknowledgmentsI thank Paul Dourish, my graduate advisor and collaborator on

the Soylent work, and my collaborators on the Roles project,

Marc Smith, Ted Welser, and Tammara Turner, as well as Paul

Johns, Eric Gleave and AJ Brush. IBM Research, Intel Research,

and the US National Science Foundation partially supported

the Soylent project. I also thank the anonymous reviewers for

useful insights and clarification.

References

1. D. Fisher and P. Dourish, “Social and Temporal Structures

in Everyday Collaboration,” Proc. Conf. Human Factors in

Computing Systems (CHI 2004), ACM Press, 2004, pp.

551–558.

2. T. Turner et al., “Picturing Usenet: Mapping Computer-

Mediated Collective Action,” J. Computer-Mediated Comm.,

vol. 10, no. 4, art. 7, 2005; http://jcmc.indiana.edu/vol10/

issue4/turner.html.

3. R.S. Burt, Structural Holes: The Social Structure of Com-

petition, Harvard Univ. Press, 1995.

4. J. Tyler et al., “Email as Spectroscopy: Automated Discov-

ery of Community Structure within Organizations,” Proc.

Communities and Technologies, Kluwer, 2003, pp. 81–96;

http://portal.acm.org/citation.cfm?id=966268.

5. F. Viegas et al., “Digital Artifacts for Remembering and

Storytelling: Posthistory and Social Network Fragments,”

Proc. 37th Hawaii Int’l Conf. System Sciences, IEEE Press,

2004, p. 10,409a.

6. B. Nardi et al., “ContactMap: Integrating Communication

and Information Through Visualizing Personal Social Net-

works,” Comm. ACM, vol. 45, no. 4, 2002, pp. 89–95.

7. S. Farnham et al., “Personal Map: Automatically Modeling

the User’s Online Social Network,” Proc. Interact 2003, IOS

Press, 2003; www.interact2003.org/docs/INTERACT03

flyer1.pdf.

8. M. Girvan and M. Newman, “Community Structure in

Social and Biological Networks, Proc. Nat’l Academy of

Science USA, vol. 98, US Nat’l Academy of Science, 2002,

pp. 7821–7826.

9. M. Smith et al., “Social Network Visualizations Of Usenet

Newsgroups,” Proc. Int’l Network Social Network Analysts

(INSNA) Conf. Social Networks (Sunbelt 2005), to be pub-

lished in 2005.

10. M. Smith, “Tools for Navigating Large Social Cyberspaces,”

Comm. ACM, vol. 45, no. 4, 2002, pp. 51–55.

11. L. Sproull and S. Kiesler, Connections: New Ways of Work-

ing in the Networked Organization, MIT Press, 1991.

Danyel Fisher is a researcher in Microsoft Research’s Commu-

nity and Technologies group. His work centers on applying

social network analysis and information visualization to

online social environments. Fisher has an MS from the Uni-

versity of California, Berkeley, and a PhD from the Univer-

sity of California, Irvine, both in computer science. He is a

member of the ACM. Contact him at danyelf@acm.org.

28 SEPTEMBER • OCTOBER 2005 www.computer.org/internet/ IEEE INTERNET COMPUTING

Social Networking

Network Visualizations

Perhaps the most recognizable featureof many social-network studies are the

ubiquitous network-wide visualizations, orsociograms.1 In general, they try to repre-sent how “far apart” nodes are by placingthem at distances proportionate to theirconnectivity with minimal distortion.Thesevisualizations can be confusing and crowd-ed when the network has many nodes orinterconnections; any distortion can makethe diagram deceptive.

Despite these concerns, the work Ipresent has leaned heavily on networkdiagrams as an interpretive tool forunderstanding social networks. With the

Soylent and Roles projects, I try to sim-plify the graphs to reasonable sizes byfocusing the graph only on single individ-uals and the persons immediately aroundthem. These graphs are far simpler, andless prone to distortion, than graphs thattry to provide overviews of broaderspaces.

Additionally, network diagrams aren’tend-user tools in my projects. Becausethey present information nonlinearly,readers can find information only bysearching entire graphics rather thanscanning simple lists. Instead, I use thegraphs as intermediate, interpretive visu-

alizations. Knowing the networks’ gener-al shapes and properties — the common-alities and differences that occur betweendifferent networks — lets developersdesign interfaces that reflect this under-lying data. I generated all visualizationswith the Java Universal Network/Graph(JUNG) toolkit.2

References1. L.C.Freeman,“Visualizing Social Networks,” J. Social

Structure, vol. 1, no. 1, 2000; www.cmu.edu/joss/.

2. J.O’Madadhain et al.,“Analysis and Visualization of

Network Data Using JUNG,” to be published in J.

Statistical Software, 2005.

Writers

For detailed information onsubmitting articles, visit www.computer.org/internet/author.htm.

Letters to the Editors

Send letters to lead editorRebecca L. Deuel at rdeuel@computer.org. Please referencethe article and issue to whichyou are responding.

On the Web

Access www.computer.org/internet/ or IEEE DistributedSystems Online athttp://dsonline.computer.org.

Changes of Address

Send change-of-addressrequests for magazinesubscriptions toaddress.change@ieee.org. Be

sure to specify IEEE InternetComputing.

Subscribe

Visit www.computer.org/subscribe/.

Article Reprints

For pricing or to place anorder, contactinternet@computer.org.

Missing or Damaged Copies

If you’re missing an issue orreceived a damaged copy,please contact membership@computer.org.

Reuse Permission

For permission to reprint anarticle, contact William Hagen,IEEE Copyrights andTrademarks Manager, atcopyrights@ieee.org.

www.computer.org/internet/

How to Reach IC