Expertise diversity of teams predicts originality and long-term impact in science and technology Weihua Li

2025-04-27 0 0 981.67KB 15 页 10玖币

侵权投诉

Expertise diversity of teams predicts originality and long-term impact in science and

technology

Weihua Li

LMIB, NLSDE, BDBC, and School of Artiﬁcial Intelligence, Beihang University, Beijing, China

Department of Advanced Interdisciplinary Research, Pengcheng Laboratory, Shenzhen, China

Zhongguancun Laboratory, Beijing, China and

Qianyuan Laboratory, Hangzhou, China

Hongwei Zheng∗

Beijing Academy of Blockchain and Edge Computing, Beijing, China

(Dated: March 25, 2025)

Despite the growing importance of collaboration networks in producing innovative science and

technology, it remains unclear how expertise diversity among team members relates to the original-

ity and impact of the work they produce. Here, drawing on statistical physics, we develop a new

computational method to quantify the expertise distance of researchers based on their prior career

histories and apply it to 23 million scientiﬁc publications and 4 million patents. We ﬁnd that across

science and technology, teams with expertise diversity tend to produce work with greater original-

ity. Teams with more diverse expertise exhibit substantially higher long-term impact (10 years),

increasingly attracting larger cross-disciplinary inﬂuence. This impact premium of expertise diver-

sity among team members becomes especially pronounced when other dimensions of team diversity

are missing, as teams within the same institution or country appear to disproportionately reap the

beneﬁts of expertise diversity. While gender-diverse teams have relatively higher impact on average,

teams with varied levels of gender diversity all seem to beneﬁt from increased expertise diversity.

Given the growing knowledge demands on individual researchers, implementation of incentives for

innovative research, and the tradeoﬀs between short-term and long-term impacts, these results may

have implications for funding, assembling, and retaining teams with originality and long-lasting

impacts.

Keywords: Collaboration networks |Diversity |Science of science |Innovation

I. INTRODUCTION

People are increasingly more likely to form collaboration networks in producing innovative work across a wide range

of creative domains[1–7]. The accumulation of knowledge[8] and the specialization of individual researchers[9, 10] un-

derscore the necessity of forming interdisciplinary teams to tackle complex challenges. In an era where narrow expertise

from a single ﬁeld may prove insuﬃcient to address pressing societal issues, collaborative eﬀorts of teams spanning

traditional disciplinary boundaries become indispensable[11–15]. Teamwork also allows researchers to establish profes-

sional connections, cultivate networks, and engage with a broader audience, both within and beyond their immediate

research community[16]. As growing scholarly attention has been paid to increasing diversity, equity, and inclusion

in science, the scientiﬁc community demands more eﬀorts to improve representational diversity in the composition of

research teams[17, 18]. A better understanding of how to assemble teams with diverse expertise is therefore essential

for coordinating collective actions, fostering interdisciplinary thinking, and integrating existing expertise to tackle

new challenges across scientiﬁc and technological domains[19–25].

However, it remains unclear how prior expertise diversity among team members is related to the originality and

the impact of the work the team produces. Some indicators have focused on quantifying the diversity of the prior

expertise among individuals, by calculating for example the Pearson correlation of the distribution across technological

classes distributions[26], a research overlap score using medical subject headings (MeSH) terms[21], a frequency-

inverse document frequency (TF-IDF) method for research content similarity[27], latent semantic analysis for the

topic similarity between researchers[24], a cosine distance estimate for prior expertise disparity[28], and a distance

metric based on the Jaccard dissimilarity of researchers’ references[29]. Although these methods can approximate

expertise diversity among team members, an eﬀective measure should also explicitly account for the relatedness of

research disciplines[30, 31]. For instance, ecology has more interactions with environmental science and evolutionary

biology than condensed matter physics. Similarly, artiﬁcial intelligence is more deeply inﬂuenced by mathematics

∗Corresponding author: hwzheng@pku.edu.cn

arXiv:2210.04422v3 [physics.soc-ph] 22 Mar 2025

and computer science compared to organic chemistry. Therefore, ecologists should on average have higher expertise

similarity to environmental scientists and evolutionary biologists than condensed matter physicists, and the knowledge

possessed by artiﬁcial intelligence researchers may be more related to the knowledge of mathematicians and computer

scientists than that of organic chemists.

In part to tackle these challenges, researchers have developed several methods to infer the diversity of knowledge

scope from the product the team has developed, by using a paper’s references or citations, using measurements such

as the distinction of ﬁelds, entropy, Gini coeﬃcient, and the Rao-Stirling (RS) index, to quantify interdisciplinarity

of the work and its association with impact[30, 32–36]. While these measures allow researchers to quantify the

interdisciplinarity of the work and its association with originality and impact, as proxies for knowledge diversity, they

depend upon the paper that a team has produced, by analyzing its references or citations, which are only available

after the fact, i.e., after the paper has been published. Moreover, these metrics are designed to quantify the diversity of

knowledge scope implemented in a speciﬁc piece of work, rather than a comparative measure to estimate the disparity

of expertise among individual scholars. It remains unclear how the composition of team expertise is related to the

originality and impact of research a team is about to produce. Understanding the association between the expertise

diversity among team members and the outcome the team produces is crucial for funding and investment decisions,

and highlights the importance of developing measures to quantify and combine diverse prior knowledge to inform

fruitful collaboration strategies.

To address these challenges, we propose a new metric to identify and quantify the diversity of prior expertise

among team members that extends beyond single disciplines. Our expertise distance metric explicitly accounts for the

relatedness of scientiﬁc ﬁelds and draws on the disciplinary distributions of prior career histories among collaborators.

The expertise diversity of a team is then obtained as the average distance of all possible pairwise coauthorships, which

correlates with a broad range of indicators of scholarly diversity in terms of the combination of past knowledge, and

other dimensions of diversity among team members regarding their aﬃliations, nationality, and gender.

A large body of work has focused on understanding the interplay between team composition and outcomes, probing

dimensions of team diversity around aﬃliations[2], ethnicity[37], gender[38], expertise[12], technical background[39],

problem-solving ability[40], intelligence[41], and more. In the technology sector, inventors from distinct social groups

tend to generate patents with greater collaborative creativity[42]. In the online knowledge-sharing community, po-

larized teams composed of a balanced proportion of ideologically diverse Wikipedia editors produce articles of higher

quality than homogeneous teams[43]. Some scholars have found that multidisciplinarity exhibits an inverted U-

shape relationship with scientiﬁc impact, although the vast majority of papers indicate a positive correlation with

multisciplinarity[44]. Furthermore, the value of multidisciplinary research might require a longer period to be rec-

ognized by the scientiﬁc community, leading to delayed impact[45]. Overall, studies from diverse domains have

demonstrated that the impact of team outcomes improves with team diversity. Thus, creative output in science and

technology may be heavily rooted in the composition of team members’ expertise and background, which is largely

determined during the team assembly process.

Originality is often regarded as a core goal in science and technology. Recent work has suggested that creative ideas

often emerge from new and unconventional combinations of knowledge from diverse disciplines, research methods, or

frameworks[16, 46]. Innovation can be spurred when proven methodologies in one domain are introduced to solve

problems in a fresh area[47]. Creativity is more likely to emerge from diverse teams as researchers integrate concepts

and methods drawn from diverse disciplines, forging connections between seemingly disparate concepts or bodies

of knowledge[1, 48]. Although studies have suggested that high multidisciplinarity is associated with high impact,

our ﬁndings suggest that scientiﬁc teamwork with multidisciplinary approaches has little correlation with originality,

quantiﬁed by the disruption score, and has negative correlations with originality in technology. Research by expertise-

diverse teams, in contrast, is positively correlated with high originality in science and technology. Therefore, despite

its close relation with multidisciplinarity, the expertise diversity of teams not only presents a new quantitative measure

to understand science but also provides new perspectives for the originality of team outcomes.

Moreover, while scholars have argued that multidisciplinarity is associated with the impact of research[49, 50], we

ﬁnd that research teams with high expertise diversity exhibit no signiﬁcant impact advantage in the short- (2 years)

or mid-term (5 years). This pattern persists for teams spanning both scientiﬁc and technological domains. We ﬁnd

that, instead, teams with high expertise diversity enjoy a substantive impact premium of their work in the long-term

(10 years), increasingly attracting cross-disciplinary inﬂuence in the longer run. The long-term eﬀect of expertise

diversity becomes more prominent as team size and citation time window grow. In particular, when other dimensions

of diversity are missing, teams formed in the same institution or country disproportionately harness the beneﬁt of

expertise diversity. These results may have implications for fostering and retaining innovative teams with more diverse

knowledge composition among team members.

II. EXPERTISE DISTANCE METRIC

We use the ﬁeld distribution of the publication history of individual authors to quantify the expertise distance

among coauthors. Suppose for ﬁeld f, the cumulative reference vector vf= (v1

f, ..., vn

f) records the numbers of

references of papers from ﬁeld fmade to all ﬁelds, where nis the number of ﬁelds[15, 30]. We then deﬁne the unit

vector of ﬁeld fas ef=vf/||vf||, where ||vf|| =qPm(vm

f)2.

For an author i, let the publication vector pirecord the number of publications author ihad in each ﬁeld up to

year t. The goal is to introduce a deﬁnition of expertise distance between a pair of authors that accounts for the

relatedness between ﬁelds. An explicit form of piis pi=a1

ie1+... +af

ief+... +an

ien, where af

iis the number

of papers author ipublished in ﬁeld f. We want to ﬁrst normalize piand obtain expertise vector qiof unit length

qi=pi/||pi|| =q1

ie1+... +qf

ief+... +qn

ienand qi= (q1

i, ..., qf

i, ..., qn

i). Note that diﬀerent from the deﬁnition of

ﬁeld vector length, here we explicitly account for the heterogeneous relatedness between ﬁelds and deﬁne

||pi||2= (a1

ie1+... +af

ief+... +an

ien)2=X

j,k

iak

i(ej·ek).(1)

Similarly, the cosine distance between expertise vectors of authors iand jcan be calculated as qi·qj=qiM qjT,

and the distance metric can be obtained as

dij =q||qi−qj||2=p2−2qi·qj=q2−2qiM qjT,(2)

where M= (eiej)i,j is the matrix indicating the closeness between ﬁelds using cosine distance. Analogously, if we

let pi= (p1

i, ..., pf

i, ..., pn

i), the length of publication vector of author ican be obtained as

||pi|| =qpiM piT.(3)

To estimate the prior expertise distance of a pair of authors (i, j) that coauthored a paper sat year t, we ﬁrst

build their respective publication vectors using all papers up to year t. We use the level 1 ﬁeld classiﬁcation in the

MAG dataset, where each article is assigned to at least one scientiﬁc ﬁeld. For the crude publication vector pi(t)

of author i,af

idenotes the number of papers ipublished in the ﬁeld fup to t. To account for authors with a

reasonably long publishing career, we include only productive authors who have published at least 5 papers up to

t. The mean expertise distance of the paper ⟨ds⟩is the average distance of all possible coauthor pairs (i, j) among

selected productive authors.

Papers with at least two productive authors are eligible to obtain a measure of team expertise distance, and those

written by solitary authors are not considered in this study. To allow for a meaningfully long prior publication record

of individual researchers, we consider only authors who have published at least 5 papers by the time of collaboration

when estimating the expertise distance of the team. Thus, a proportion of less productive or early-career researchers

are dropped and we obtain a subset of team-authored papers eligible for a distance metric.

III. DATA

We use the publication and citation data in 1950-2019 from the Microsoft Academic Graph (MAG) dataset[51]. We

select papers published in journals for science, technology, and social sciences, namely biology, business, chemistry,

computer science, economics, engineering, environmental science, geography, geology, materials science, mathematics,

medicine, physics, political science, sociology, and papers published in conferences for computer science. We also

extract patent data from the MAG archive.

There are some known limitations and deﬁciencies of MAG, particularly regarding citation data[52]. To address

this issue, we implement a rigorous data ﬁltering process before the analysis (see also Supplementary Information).

In particular, we exclusively include papers sourced from journals or conference proceedings within MAG, and ensure

that all papers contain author aﬃliation information. After data ﬁltering procedures, we retain a total number of

22.8 million papers, and 354 million citations from papers to papers. We also extract 4.4 million patents, which make

29.9 million citations among themselves, and 5.6 million citations to the selected subset of research articles.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Expertisediversityofteamspredictsoriginalityandlong-termimpactinscienceandtechnologyWeihuaLiLMIB,NLSDE,BDBC,andSchoolofArtificialIntelligence,BeihangUniversity,Beijing,ChinaDepartmentofAdvancedInterdisciplinaryResearch,PengchengLaboratory,Shenzhen,ChinaZhongguancunLaboratory,Beijing,ChinaandQianyuan...

展开>> 收起<<

Expertise diversity of teams predicts originality and long-term impact in science and technology Weihua Li.pdf

共15页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Expertise diversity of teams predicts originality and long-term impact in science and technology Weihua Li

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: