BLUEPRINTS FOR SUCCESS

Guidelines for Building Multidisciplinary Collaboration Teams

Sidath Gunawardena and Rosina O. Weber

The iSchool, Drexel University, Philadelphia, Pennsylvania, U.S.A.

Keywords: Recommender Systems, Case-Based Reasoning, Multidisciplinary Collaboration.

Abstract: Finding collaborators to engage in academic research is a challenging task, especially when the

collaboration is multidisciplinary in nature and collaborators are needed from different disciplines. This

paper uses evidence of successful multidisciplinary collaborations, funded proposals, in a novel way: as an

input for a method of recommendation of multidisciplinary collaboration teams. We attempt to answer two

questions posed by a collaboration seeker: what disciplines provide collaboration opportunities and what

combinations of characteristics of collaborators have been successful in the past? We describe a two-step

recommendation framework where the first step recommends potential disciplines with collaboration

potential based on current trends in funding. The second step recommends characteristics for a collaboration

team that are consistent with past instances of successful collaborations. We examine how this information

source can be used in a case-based recommender system and present a preliminary validation of the system

using statistical methods.

1 INTRODUCTION

Multidisciplinary collaboration brings together

groups of researchers from different fields to solve a

common problem, one that cannot be solved using

the theories and methods of a single field (National

Academies, 2005). US federal agencies encourage

multidisciplinary research through increased funding

initiatives (National Academies, 2005; National

Science Foundation, 2006). Obtaining such funding

is one way that academics, particularly tenure-track

junior faculty, can advance their careers (Higgins

and Walsh, 2009). Thus, academic researchers may

need to find collaborators in fields very different

from their own.

The traditional methods for finding a

collaborator, such as leveraging one’s professional

ties, attending conferences, joining learned societies,

and participating in on-line discussion groups

(Clegg, 2003), by their nature, tend to focus inwards,

towards one’s own discipline (Kogan, 2000). Thus,

such methods are much more likely to be successful

when employed to find a collaborator in one’s own

discipline than when used to find a partner in a

different discipline. Junior faculty members are at

even greater disadvantage as they lack both

experience and personal ties.

Currently available technological means provide

little assistance in solving this problem.

Technologies that leverage social networks to

identify collaborators are limited to single

disciplines (Ayanegui-Santiago et al., 2009; Liben-

Nowell & Kleinberg, 2003; Newman, 2001). Expert

locator systems focus on either finding an individual

with pre-specified expertise or an expert able to

answer to a pre-specified question (Serdyukov et al.,

2008). They solve a very narrow problem of locating

an expert to meet a pre-specified short term

knowledge need. Hence, there is scope for a

systematic, technological method for recommending

synergistic disciplines and the desired characteristics

of potential collaborators.

In order to find data that can help provide useful

guidance, we look to existing successful

multidisciplinary collaborations. In the context of

competitive grant funding, we find repositories of

experiences of successful multidisciplinary

collaborations in the form of funded grant proposals.

In order to make proper use of those experiences, we

adopt a Case-Based Reasoning (CBR) methodology,

a reasoning methodology that enables the reuse of

experiences in multiple forms (Bride et al., 2005).

While recommender systems are found in myriad

contexts, we have yet to find any that attempt the

task of recommending collaborators for

387

Gunawardena S. and O. Weber R..

BLUEPRINTS FOR SUCCESS - Guidelines for Building Multidisciplinary Collaboration Teams.

DOI: 10.5220/0003753503870393

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 387-393

ISBN: 978-989-8425-95-9

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

multidisciplinary research.

In the next section we present some background

literature, we then detail our data sources in Section

3. In Section 4 we present our methodology and in

Section 5 our experiments and a discussion of our

results. We close with our conclusions, and some

thoughts about future work.

2 BACKGROUND

Recommending multidisciplinary collaborations has

not been explored before, so the background of this

work comes from recommending collaborators

within the same discipline and also at the work on

locating experts

2.1 Social Networks

The links between researchers created by co-

authorship, co-publication, or citation, can be

leveraged to create social networks (Barabási et al,

2002; Tang et al., 2008), with co-authorship being

the strongest link. In the case of co-authorship, the

‘distance’ between two authors is represented by the

number of links that have to be traversed to make

the connection between them. The number of co-

authorships between two authors can be used as a

measure of the strength of such linkages (Newman,

2001). Social networks can also be combined with

other approaches as expert location systems to

improve their usefulness to users by taking into

account social dynamics in addition to expertise

(McDonald, 2003). Work in social networking

shows some promise for discovering collaborators

who have the potential to work together, but the

work is limited to researchers in the same field

(Ayanegui-Santiago et al., 2009; Newman, 2003).

2.2 Expert Locator Systems

Collaborator recommendation is related to expert

locator systems (ELS) (Becerra-Fernandez, 2003);

where the system can recommend qualified experts

to a user who has a need for a particular expertise.

The level of expertise must be narrowly defined

either as a question that needs an expert answer

(Serdyukov, 2008) or limited to one organization

(Maybury, 2002; McDonald 2003). When the user

needs a particular type of expertise, the system

selects the candidate that best matches the user’s

expertise criteria. Additional factors such as

availability can also be taken into account

(McDonald & Ackerman, 2000).

When seeking a collaborator, the criteria to be

satisfied are vague and ill-defined. We define

researchers seeking to engage in multidisciplinary

collaboration as collaboration seekers. The

collaboration seeker likely does not know all the

domains where suitable collaboration partners

reside. Furthermore, factors additional to expertise

need to be included. Hence, we perceive the

potential usefulness of recommender systems.

We see collaboration recommendation and

expert location as two separate parts of the process

of finding a collaborator. The recommendation

identifies the disciplines and the characteristics of

the collaborators, and subsequently, expert location

is used to identify the specific individuals who meet

those characteristics.

2.3 Collaboration

A summary of some of the literature on

collaboration can be found in Gunawardena et al. (

2010). Collaboration is an idiosyncratic process, and

when it occurs across disciplinary boundaries it can

create or exacerbate issues such as trust, the need for

negotiation, and the need for a common vocabulary

(Jeffrey, 2003). Thus, when recommending

collaboration teams, factors that can mitigate such

problems need to be taken into account.

Collaborators who are nearby and can facilitate face

to face communications (Kat, 1994), senior

colleagues can act as mediators (Bozeman & Corley,

2004; Wood & Gray, 1991), and collaborating with

those at institutions with high research productivity

can be beneficial (Jones et al., 2008). We examine

data sources to find reasonable proxies for these

factors. An initial experiment on this problem used

funded grants but was limited to only area of

expertise (Gunawardena & Weber 2009) showed

that even with limited information it was possible to

provide a basic recommendation. This work

broadens the scope to include additional features of

researchers known in the literature to have an impact

on collaborative behaviors: the researchers’ location,

their title, which is used as a proxy for their

seniority, and the type of institution they belong. We

take the most literal definition of multidisciplinary;

in the collaborations we study are required to

contain at least two members who have different

departmental affiliations.

2.4 Case-based Reasoning

In CBR, the cases are typically composed of a

problem context and a lesson that can be learned

ICAART 2012 - International Conference on Agents and Artificial Intelligence

388

www.proquest.com

www.academicanalytics.com

about it (Kolodner, 1993). The lesson can be thought

of as the as the solution applicable to that particular

problem context. In a case-based recommender

system this takes the form of collection (case-base)

of problems and associated solutions. A new

problem is solved by reusing the solution of the most

similar old problem (Bridge et al., 2005). We

approach the problem of recommending

collaborators by looking at what lessons we can

learn from past successful collaborations.

In collaboration recommendation the problem to

be solved is finding suitable collaboration partners

for a faculty member, who is described by a set of

characteristics (title, research area, institution, etc).

The solution is described by the characteristics of

the faculty with the best potential for collaborative

success. Here the solution is presented by the same

features that are used to describe the problem. Thus,

the process of recommendation for a new

collaboration seeker involves searching the case

base for the collaboration with a member most

similar to the collaboration seeker and then

recommending the remaining collaborators in that

collaboration, that is, the complementary portion of

the collaboration, as the recommended collaboration

team.

3 DATA SOURCES

We use funded grant proposals as experiences of

successful multidisciplinary collaborations. The

grant proposals contain the name and affiliated

institution of the principal investigator and the

names of the co-investigators. Thus, the information

pertaining is incomplete with respect to what is

required for solving the recommendation problem.

To obtain a fuller picture of the collaborations we

use additional sources of information.

3.1 Grant Data

For our experiments, we use grants funded by the

Office of Multidisciplinary Activities (OMA), a

directorate of the National Science Foundation

(NSF), whose goal is to fund research in the

mathematical and physical sciences that crosses

disciplinary boundaries. We also utilize two

additional sources to obtain the data required for

these experiments. COS Scholar Universe, is a

database of 2 million profiles of full time faculty

supported by ProQuest LLC

. We obtain our data on

researchers’ departmental affiliations and titles from

this source. Our third source of data is Academic

Analytics LLC

, a private company that provides the

ranking of doctoral programs. We obtain our

information on institution type and location from

this source.

3.2 The Data Set

The dataset includes NSF grants from the period

2005-2010 that are composed of two to five

members, with at least two members from different

departments. The dataset contains 173

collaborations, involving 530 total faculty members

from US academic institutions.

We aggregated the data, limiting the

collaborations chosen to those comprised only of

researchers with the titles of Assistant Professor,

Associate Professor and Full Professors. Table 1

presents a summary of the data, and how it is coded.

The departmental names have non-relevant terms

removed to assign values to the feature Discipline

(e.g. Department of Physics would be reduced to

Physics).

Table 1: Summary of data.

Feature Description

Title Full, Associate, or Assistant Professor

Discipline

143 possible values (Chemistry, Astrophysics,

Civil Engineering, …)

Institution

Type

Large Research Inst, Small Research Inst,

Specialized Inst.

Institution

Location

Region (Northeast, Midwest, South, West)

We use the definition employed by Academic

Analytics to categorize institutions by type. A

university is considered a Large Research University

(LRU) if it has at least fifteen PhD programs each

with at least ten faculty members. A Small Research

University (SRU) has between one and fourteen PhD

programs. A Specialized University is one that

awards a majority of their degrees in one field.

4 METHODOLOGY

In this section we describe the evolution of our

research process, as we sequentially developed our

method, with each step of the process informing the

design of the subsequent experiments.

BLUEPRINTS FOR SUCCESS - Guidelines for Building Multidisciplinary Collaboration Teams

389

4.1 Similarity Functions

We begin by explaining similarity in CBR and go

onto describe the similarity functions we employ.

In CBR, the similarity function determines

which cases in the case-base are selected, and thus

which solutions are reused. The similarity function

compares the characteristics of the new problem to

the problems in the case-base and gives each case a

score based on how similar it is to the new problem,

with the higher scores assigned to the candidates to

have their solutions reused.

Our initial analyses employed standard

similarity methods: weighted and unweighted

feature counting. We compared these to a baseline

method of random recommendation and also to a

modified random recommendation based on

location. The purpose of the experiments is to

demonstrate that the data does contain knowledge to

make recommendations and then build on that to

determine how to make more accurate

recommendations.

4.1.1 Baseline Method: Random

Recommendation

A collaborator is selected from the dataset and then

n collaboration teams are randomly selected, with no

team being selected twice, where n has the set of

values {1, 3, 5, 10}.

4.1.2 Random Recommendation by Located

Region

A collaborator is selected from the dataset and then

randomly n collaboration teams are selected from

the same region as the original collaborator, with no

team being selected twice, n {1, 3, 5, 10}.

4.1.3 Feature Counting

As a first step, this method considers the selected

features to have equal importance for similarity

assessment. In a feature counting method, the

similarity between the target artificial case t and

candidate case c is given by Equation (1):

Similarity =

.Sim(t



)





(1)

Where n is the number of features and Sim(t

) = 1

if t

= c

, and 0 otherwise. Each collaboration has as

many candidates as members. The similarity score

used is the highest score obtained from all members.

The remaining collaborators in that collaboration

will be the team that is recommended.

4.1.4 Weighted

The weighted similarity method takes into

consideration the relative importance of the features.

Here the similarity between the target artificial case t

and candidate case c is given by Equation (2):

Similarit

.w



Sim(t



)





(2)

Where n is the number of features, w



is the weight

associated with feature i, and Sim(t

) = 1 if t

= c

and 0 otherwise.

To determine weights, we employ a genetic

algorithm, a machine learning method used for

optimization. It is based around the evolutionary

principle of survival of the fittest, that is, in a

population, the strongest genetic chromosomes

survive and are passed on to future generations

(Kelly & Davis, 1991). Genetic algorithms are a

common method to derive weights for use in CBR

systems (Beddoe & Petrovic, 2006; Dogan et al.,

2006; Fu & Shen, 2004; Jarmulak et al., 2000). In

this experiment, each characteristic of a collaborator

(title, research interest, etc) is a chromosome. A

genetic algorithm can be broken down into the

following steps: initial weight generation, fitness

evaluation, reproduction (including possible

mutation). It also requires a predefined stopping

criterion to terminate the process. For this

experiment we apply a genetic algorithm with the

following parameters: a crossover of 0.5 where each

parent has an equal chance of providing the

chromosome, a 1% chance of mutation where a gene

is replaced by a new, random chromosome. The

fitness function which determines which genes go to

the next generation is determined based on accuracy

at the top1 threshold. The algorithm will stop after

100 iterations. The execution of the genetic

algorithm produced the following weights:

Table 2: Genetic algorithm derived weights.

Title Discipline Region Inst. Type

0.24 0.34 0.34 0.08

4.2 Two Step Recommendation

There are two broad dimensions required to be

considered when making this particular

recommendation: a collaborator’s research interest

and their personal characteristics. The derived

weights suggest that, combined, the personal

ICAART 2012 - International Conference on Agents and Artificial Intelligence

390

http://www.nationalacademies.org/

characteristics (title, region, institution type)

combined have a greater importance than that of

research interest. This does not make intuitive sense

as if a mathematician is seeking to engage in

collaboration, then the previous collaborations of,

for example, biologists have little value for the

purposes of identifying potential domains. Thus, we

take into account the practical aspects of a useful

recommendation, similar to Baccigalupo & Plaza

(2007) who in their work on song recommendation

ignore songs that are irrelevant based on the user’s

specifications. Here the discipline is the primary

determining factor, and the other factors secondary.

To reflect this, in this experiment, we break the

recommendation process into two steps.

Step 1: determine all the cases in the case-base

that could provide useful recommendations. This is

done by limiting the cases used to those that have at

least one member from the same discipline or a

discipline that is a sibling on a disciplinary

taxonomy as the collaboration seeker. For our

experiments we use the taxonomy used by the

National Academies to classify doctoral programs

Step 2: recommend the secondary

characteristics of collaborators based on the

characteristics of the collaborations seeker. We use

the remaining features, title, location, and institution

type to then recommend a potential team: the

complementary portion of that collaboration.

The recommendation of the disciplines is

decoupled from the recommendation of the

characteristics of collaborators. Thus, with the two

step approach the system is no longer limited to

recommendations that exist as collaborations within

the case-base. It can recommend the disciplines from

one collaboration with the collaborator

characteristics of another if it determines that that is

the best recommendation for a particular

collaboration seeker.

4.2.1 Feature Counting with Two-step

In the first step we limit the cases to those that have

at least one member from the same discipline or a

discipline that is a sibling on the disciplinary

taxonomy as the collaboration seeker. Then we

perform the feature counting similarity assessment

as before, but only using title, location and

institution type as features.

4.2.2 Weighted with Two-step

Here too we apply the two-step approach, using the

first step to reduce the case-base and then run the

GA to determine the weights of the remaining three

features. We execute the GA using the same

parameters as before. The execution of the genetic

algorithm produced the following weights:

Table 3: Genetic algorithm derived weights.

Title Region Inst. Type

0.26 0.51 0.23

Thus we have the following hypotheses:

H1: Randomly selecting teams by region is

more accurate than random selection.

H2: The feature counting method is more

accurate than randomly selecting teams by

region.

H3: The weighted method is more accurate than

the feature counting method.

H4: The 2 step feature counting method is more

accurate than the feature counting method.

H5: The 2 step weighted is more accurate than

the weighted method.

H6: The 2 step weighted is more accurate than

the 2 step feature counting method.

5 EXPERIMENTS

In this section we present the experiments we

conducted on the grant dataset to demonstrate the

effectiveness of this approach. These experiments

are used to increase our understanding of the data, to

allow us to determine whether it can be utilized to

make useful recommendations.

5.1 Evaluation

A leave-one-out cross-validation (LOOCV) is a

standard method to evaluate recommender systems.

To apply LOOCV, a collaboration is removed from

the collection and its members used as target cases.

Accuracy is measured by whether the system

retrieves the most similar case to the complementary

portion of the removed case. However, we do not

have the ability to determine similarity between

collaborations to determine second best solution. To

overcome this hurdle, we use what we term

‘artificial collaboration seekers’ who we can

artificially create as being very similar to the

original collaborators in the system. We describe

this process in the following section.

BLUEPRINTS FOR SUCCESS - Guidelines for Building Multidisciplinary Collaboration Teams

391

5.2 Generating Artificial Collaboration

Seekers

From a collaboration we select each collaborator in

turn and randomly select one of the features

(discipline, title, institution type, or location) and

modify it. The modification is such that when a

feature value is modified, it is changed to an

adjacent value, that is, a collaborator’s title may

change from assistant to associate professor, but not

to full, where as an associate professor may be

changed to either a full professor or an assistant

professor. If the feature to be modified is discipline,

then we use the taxonomy and modify the discipline

and replace it with one that is a sibling.

5.3 Accuracy

In our experiments we measure accuracy as follows:

when an artificial collaboration seeker is submitted

to the system as a new target problem the retrieval

set contains the complementary members of the

original collaboration that generated the artificial

collaboration seeker is retrieved within the top n

cases. We examine results for the top n cases,

considering n = {1, 3, 5, 10}. Tied values are

considered to be equivalent in rank when

determining whether a particular retrieval was

successful or not. An artificial collaboration seeker

is created for every collaborator in the dataset and

accuracy is measured by whether collaboration team

of the original collaborator is one of the top n

recommended teams, n = {1, 3, 5, 10}. Each

experiment is repeated ten times with an average

accuracy calculated.

A one-way ANOVA test is used to determine if

there is a significant difference between the means

of the various methods (α = 0.05), post hoc analyses

of Scheffe, Tukey’s Honest Significant Differences,

Bonferroni Adjustment, and Least Significant

Differences are then used to perform multiple

comparisons between the means. A difference is

reported as significant only if all four tests concur.

The random methods are outperformed at all levels

of accuracy, but the other methods only show a

significant difference only when the top result is

considered.

5.4 Results and Discussion

Based on the post hoc analysis at the 0.05

confidence level, we are able to reject the null

hypothesis that there is no difference between the

random methods and the feature counting and

weighted methods at all levels of accuracy. In

addition, at the top level of accuracy, the weighted

methods outperform their feature counting

counterparts and the two-step method shows an

improvement in accuracy in both weighted and

feature counting methods (Table 4). No significant

difference was observed between these 4 methods at

other levels of accuracy.

Table 4: Average accuracy (standard deviation) top1

results.

Feature Counting 0.492 (0.012)

2Step Feature Counting 0.521 (0.012)

Weighted 0.526 (0.017)

2Step Weighted 0.551 (0.011)

Our results versus a random baseline show that this

data does possess knowledge and can be used as the

basis for the recommendation of multidisciplinary

collaboration teams. The subsequent results are

mixed, showing statistically significant improvement

only at the top level of accuracy. This is less

improvement than expected of the two step method.

However, the two step method recommends the best

potential collaboration, which may not be one that

exists in the case-base, penalizing its accuracy.

6 CONCLUSIONS AND FUTURE

WORK

In this paper we show how funded grants may be

used as a basis for solving a novel problem:

recommending multidisciplinary collaboration

teams. Using the grant dataset, we demonstrated that

the proposed approach can provide

recommendations that are superior to random, and

showed further improvements to increase their

quality. These results suggest this is a viable

approach to using this data on this problem. This

approach has room for improvement but it is unique

in its use of the data and in providing a solution to

this problem. Out of many possible improvements,

we name a few. Instead of discipline the use of

publication keywords can provide a more detailed

recommendation. Additionally, these experiments

focus solely on analogical reasoning, incorporating

analytical knowledge from the literature on

collaboration may add to the quality of the

recommendation.

REFERENCES

Aha, D. (1997) Editorial on Lazy Learning. Artificial

ICAART 2012 - International Conference on Agents and Artificial Intelligence

392

Intelligence Review, 11, 7-10

Ayanegui-Santiago, H., Reyes-Galaviz, O., Chávez-Ara-

gón, A., Ramírez-Cruz, F., Portilla, A., & García-Ba-

ñuelos, L. (2009). Mining Social Networks on the Me-

xican Computer Science Community. In MICAI 2009:

Advances in Artificial Intelligence (pp. 213-224).

Barabási, A. L., Jeong, H., Néda, Z., Ravasz, E., Schubert,

A., & Vicsek, T. (2002). Evolution of the social net-

work of scientific collaborations. Physica A:

Statistical Mechanics and its Applications, 311(3-4),

590-614.

Becerra-Fernandez, I. (2006). Searching for Experts on the

Web: A Review of Contemporary Expertise Locator

Systems, ACM Transactions on Internet Technology,

6(4), 333-355.

Beddoe, G. R., & Petrovic, S. (2006). Selecting and

weighting features using a genetic algorithm in a case-

based reasoning approach to personnel rostering. Eu-

ropean Journal of Operational Research, 175(2), 649-

671.

Baccigalupo, C. and Plaza, E. (2007), A Case-Based Song

Scheduler for Group Customised Radio. Proc. ICCBR

2007, Lecture Notes in Computer Science Vol. 4626,

p. 433-448. Springer Verlag.

Bozeman, B. and Corley, E. 2004. "Scientists' colla-

boration strategies: implications for scientific and

technical human capital," Research Policy, vol. 33, no.

4, pp. 599-616.

Bridge, D., Goker, M. H., McGinty, L, Smyth, B. (2005)

On Case-Based Recommender Systems. Knowledge

Engineering Review, 20 (3):315-320

Clegg, S. 2003. Problematising ourselves: Continuing pro-

fessional development in higher education. Interna-

tional Journal for Academic Development, 8, 37-50.

Dogan, S. Z., Arditi, D., Gunayd, X, & N, H. M. (2006).

Determining Attribute Weights in a CBR Model for

Early Cost Prediction of Structural Systems. Journal

of Construction Engineering and Management,

132(10), 1092-1098.

Fu, Y., & Shen, R. (2004). GA based CBR approach in

Q&A system. Expert Systems with Applications, 26(2),

167-170.

Gunawardena, S., & Weber, R. (2009). Discovering Pa-

tterns of Collaboration for Recommendation. In Pro-

ceedings 22th International FLAIRS Conference, FLA

IRS'09. AAAI Press, Menlo Park, California, 2009.

Gunawardena, S., Weber, R., & Agosto, D. E. (2010). Fin-

ding that Special Someone: Modeling Collabo-ration

in an Academic Context. Journal of Education for Li-

brary and Information Science, 51.

Higgins, S. E., & Welsh, T. S. (2009). The Tenure Process

in LIS: A Survey of LIS/IS Program Directors.

Journal of Education for Library and Information

Science, 50(3), 176-189.

Jarmulak, J., Craw, S., & Rowe, R. (2000). Genetic Al-

gorithms to Optimise CBR Retrieval. In Advances in

Case-Based Reasoning (Vol. 1898, pp. 159-194):

Springer Berlin / Heidelberg.

Jeffrey, P. (2003). Smoothing the Waters: Observations on

the Process of Cross-Disciplinary Research Colla-

boration. Social Studies of Science, 33(4), 539-562.

Jones, B. F., Wuchty, S., & Uzzi, B. (2008). Multi-

University Research Teams: Shifting Impact, Geo-

graphy, and Stratification in Science. Science,

322(5905), 1259-1262.

Katz, J. (1994). Geographical proximity and scientific co-

llaboration. Scientometrics, 31(1), 31-43.

Kelly, J. and Davis, L. (1991). A hybrid genetic algorithm

for classiﬁcation. In Proceedings of the 12th IJCAI,

Sidney, Australia, 645–650.

Kogan, M. (2000). Higher Education Communities and

Academic Identity. Higher Education Quarterly, 54,

207-216.

Kolodner, J. (1993). Case-Based Reasoning. Morgan

Kaufmann, San Francisco

Liben-Nowell, D., & Kleinberg, J. (2003). The link

prediction problem for social networks. Paper pre-

sented at the Proceedings of the twelfth international

conference on Information and knowledge

management.

McDonald, D. W., & Ackerman, M. S. (2000). Expertise

recommender: a flexible recommendation system and

architecture. Paper presented at the Proceedings of the

2000 ACM conference on Computer supported coo-

perative work.

Maybury, M. T. 2002. Knowledge on demand: Knowledge

and expert discovery, Journal of Universal Computer

Science (8)5, pp. 491-505.

McDonald, D. (2003). Recommending collaboration with

social networks: a comparative evaluation. Paper pre-

sented at the CHI '03: Proceedings of the SIGCHI

conference on Human factors in computing systems,

Ft. Lauderdale, Florida, USA.

National Academy of Sciences, National Academy of En-

gineering, Institute of Medicine. (2005). Facilitating

Interdisciplinary Research. Washington, DC.:

National Academies Press.

National Science Foundation, Strategic Plan FY 2006-

2011. Investing in America’s Future (NSF 06-48),

September 2006.

National Academy of Sciences, National Academy of En-

gineering, Institute of Medicine. (2005). Facilitating

Interdisciplinary Research. Washington, DC.:

National Academies Press.

Newman, M. E. J. (2001). From the Cover: The structure

of scientific collaboration networks (Publication no.

10.1073/pnas.021544898).

Serdyukov, P., Feng, L., van Bunningen, A., Evers, S., van

Heerde, H., Apers, P., Fokkinga, M., and Hiemstra, D.

(2008). The Right Expert at the Right Time and Place.

PAKM: 38-49.

Tang, J., Zhang, J., Yao, L., Li, J., Zhang, L., & Su, Z.

(2008). ArnetMiner: Extraction and Mining of Aca-

demic Social Networks. Paper presented at the

Fourteenth ACM SIGKDD International Conference

on Knowledge Discovery and Data Mining

(SIGKDD'2009), Las Vegas, Nevada, USA.

Wood, D., & Gray, B. (1991). Towards a Comprehensive

Theory of Collaboration. Journal of Applied

Behavioral Science, 27(2), 139-162.

BLUEPRINTS FOR SUCCESS - Guidelines for Building Multidisciplinary Collaboration Teams

393