Application of Social Network Analysis to …

1 Application of Social Network Analysis to collaborative team FormationMichelle CheathamKevin CleeremanInformation DirectorateInformation DirectorateAFRLAFRLWPAFB, OH 45433 WPAFB, OH formation is a challenging problem in many large or-ganizations in which it is entirely possible for two individ-uals to work on similar projects without realizing it. Byapplying Social Network Analysis to mappings of co-authorsand to mappings of related research paper keywords, weare able to help generate teams of diverse individuals withsimilar interests and : Social Network , concept map, collabora-tion, team INTRODUCTIONThe first step in collaboration is knowing who to collabo-rate with. But today s organizations are so fast-paced andgeographically distributed that there may be two employ-ees working on the exact same project without realizing , the company may be employing outside consultantsunnecessarily.

2 An effective way to insure that organizationsare taking advantage of their human capital is needed. Typ-ical approaches to this problem include reorganizing to putemployees working on similar projects in close geographicproximity and the use of yellow page systems. Reorga-nizations can be a disruptive shock to employees and areimpractical since they would need to be done frequently asemployees skill sets change over time. Yellow page sys-tems generally ask an employee to fill out a questionnaireabout his skill set which can then be searched by are various problems with this approach, includingkeeping the directory current and relying on employees notto be too modest or boastful when listing their skills [6]. Inthis paper we consider using a modified version of socialnetwork Analysis to construct a graph of employee interac-tions along with the subjects these employees are workingon.

3 This graph can then be used to suggest new collabora-tive Network Analysis (SNA) is a method of studying in-teractions among individuals or groups. SNA is best appliedin situations where the data is inherently relational, mean-ing it is a property of theinteractionof agents as opposedto individual agents [7]. Examples include computer net-works, organizational relationships, and family trees. Muchwork has been done based on email or IM traffic [6], [8], [3],but this type of research raises privacy concerns. Newmanpoints out the benefit of basing SNA on affiliation networks(a Network in which people are related by membership ina common group or club): the data is readily available anddoes not rely on questionnaires or interviews [5] as is thecase with yellow page systems. Plus, the information is of-ten public, eliminating privacy concerns.

4 For this project wehave used coauthorship on published works as the basis forconstructing the Social Network graph. Other types of data,such as a list of project titles and their teams or meetingminutes and the participants are also viable datasets. Thispaper first shows two different views of the paper-authordata and considers the types of questions that can be an-swered with those views. We then take the Analysis onestep further and use the keywords of the papers to developa graph of papers and the concepts they cover. Combiningthis information with the standard SNA allows us to suggestnew collaborative RELATED WORKP erforming Social Network Analysis based on coauthor-ship is not a new area of interest. In particular, New-man has specifically considered scientific collaboration net-works based on publication information in several scientificdatabases [4].

5 His work found that these networks displaythe small world characteristic, are highly clustered, andobey a power-law distribution with an exponential did not go into detail on what these characteris-tics meant in the scientific collaboration context nor did heconsider the implications for potential future collaborations, attempted to discover shared interests using graphanalysis in the early nineties [6]. His method involved ana-lyzing email traffic from 15 different organizations and run-ning various algorithms on the resulting graph to determinewhich individuals shared his own interests. While the al-gorithms developed are very interesting, this work did notexplicitly considerwhichsubjects the individuals shared aninterest in, only that they , done by Kautz s group at AT&T Laboratories,is an interesting Application in this area.

6 It mines publiclyavailable documents on the internet and constructs a graphof names that appear in close proximity [2]. The resultingsystem can then answer questions like What is my rela-tionship with Person A? and What people in my neigh-borhood know about topic x? The method that Referral-Web uses to answer this later type of question is not clear,however, and team formation is not mentioned. Khan et considered both the authors and subjects of pa-pers, but the focus was on examining existing collaborationsand predicting new ones [3]. The group did not explore theutility of their system for collaborative team TRADITIONAL SNAThe data we used consists of all the publications producedby a branch within the Air Force Research Laboratory be-tween 2003 and 2005. To preserve employee privacy, thenames of the authors have been replaced by numbers.

7 Thedataset consisted of 71 papers written by 80 different au-thors. The average number of papers written by an authorwas , and the average number of coauthors on a sin-gle paper was The average author collaborated other people. Newman s corresponding numbers for theMEDLINE, Los Alamos, and NCSTRL databases contain-ing scientific publications were , , and respec-tively. The lower numbers for our dataset are likely due tothe limited timespan considered and the omission of paperswritten by academics and contractors besides those done inconjunction with 1 and 2 show two traditional views of the Social net-work graph for this data. In Figure 1, the nodes of the graphare authors, and the lines between them indicate that theconnected authors have collaborated on a paper. The thickerthe line, the more papers they have worked together view makes it easy to see the different workgroups thatexist, as well as which individuals primarily work alone andwhich serve as bridges between different workgroups.

8 Inthe dataset examined here, there are seven major clustersand several smaller groups. It is readily observed that au-thors 63 and 13 have collaborated extensively (with one an-other and with a group of other colleagues) while 29 hasonly worked individually on publications. 13 also acts asthe only bridge between two large groups of graph also shows that 52 acts as a hub for a group ofeight researchers. This gives a coarse indication of what in-dividuals would work well together in a group setting. Forexample, creating a group that consisted entirely of employ-ees who had only worked alone on papers in the past, with-out any of the potentially more outgoing hub employees,Figure 1: Traditional Viewmay not be productive. In addition, if employees have al-ready worked together numerous times in the past they willlikely be able to function smoothly in a group from the start,whereas employees from different clusters will bring a morediverse set of viewpoints to bear on a problem at the expenseof easy communication [1].

9 Newman suggests that affiliation networks are fundamen-tally bipartite graphs with one type of node representing in-dividuals and the other representing the groups they belongto and edges can only connect vertices of unlike type [5].In Figure 2, the same dataset is shown as a bipartite graphwhere one class is the authors (represented as squares) andthe other is the papers (represented as circles). The size ofthe node is based on its degree, so that more prolific au-thors and papers with more coauthors are larger. This viewstill allows us to see the different workgroups that exist, butnow it is easier to see what those groups are working graph also makes it easy to determine who has beenpublishing the most and which papers have been the focusof attention. In the graph of our data, it is easy to see that52 has written the most papers while Real Time StreamingData Grid Applications had the most contributors (shownin black).

10 We can also use this graph (or the one shown inFigure 1) to answer some common queries in Social networkanalysis, such as How are 3 and 44 connected? (Figure3) and What neighborhood of people can be reached in agiven number of levels, starting from 51? (Figure 4). Thistype of Analysis has implications for team formation . Kautzreasons in [2] that an individual is more likely to trust aperson recommended to them if they can see the chain ofknown acquaintances between themselves and the recom-Figure 2: Bipartite ViewFigure 3: Shortest path between two authorsmended person. In addition, if there are many people in thesocial Network who cannot reach one another ( manypeople whose neighborhood is a small subset of the overallgraph), then knowledge cannot effectively spread through-out the branch or organization.