=Paper=
{{Paper
|id=Vol-3745/paper14
|storemode=property
|title=Connector and Provincial Hub Dichotomy in Scientific Collaborations Identified by Reinforcement Learning Algorithm
|pdfUrl=https://ceur-ws.org/Vol-3745/paper14.pdf
|volume=Vol-3745
|authors=Feifan Liu,Shuang Zhang,Haoxiang Xia
|dblpUrl=https://dblp.org/rec/conf/eeke/LiuZX24
}}
==Connector and Provincial Hub Dichotomy in Scientific Collaborations Identified by Reinforcement Learning Algorithm==
https://ceur-ws.org/Vol-3745/paper14.pdf
Connector and Provincial Hub Dichotomy in Scientific
Collaborations Identified by Reinforcement Learning Algorithm⋆
Feifan Liu1,2 , Shuang Zhang1 and Haoxiang Xia1,2,*
1
Institute of Systems Engineering, Dalian University of Technology, No.2 Linggong Road, Dalian, 116024, Liaoning, China
2
Institute for Advanced Intelligence, Dalian University of Technology, No.2 Linggong Road, Dalian, 116024, Liaoning, China
Abstract
Scientific problem-solving relies on effective organizational patterns of research collaboration. To recognize the more complex cross-
community collaboration patterns of researchers in modern science, this study probes the central core structure of co-authorship networks
at the mesoscale, aiming at understanding the emerging structural characteristics and functional performance of the effectiveness of
complex research and innovation systems. Taking the field of physics as an example, combining the deep reinforcement learning pre-
training model with the hub role information of the complex network, this study identifies both the provincial hub and global connector
hub and the emergence of multi-core structures at the mesoscopic level of the scientific collaboration network. The existence of the
multi-core structure reflects the spontaneous formation of "local centrality and global decentrality" in the scientific collaboration system,
which makes the knowledge creation system economical at the structural level and efficient in the functions of global collaboration and
knowledge diffusion. Through an analysis of the structural and functional characteristics and mesoscale collaborative organizational
structures of researchers, this study enhances comprehension and insights into the inherent factors propelling scientific development
and the dynamics of collective knowledge creation. The findings contribute valuable perspectives for the establishment of inclusive
scientific research management policies, fostering a more sophisticated scientific research and innovation system.
Keywords
Scientific collaborative behavior, complex network analysis, deep reinforcement learning, hub role identification
1. Introduction This study aims to address these pressing issues by
identifying and examining multi-core structures within co-
The scientific research and innovation system embodies a authorship networks using a mesoscopic lens that taps into
form of "collective intelligence," where individual scholars the inherent community structure. Leveraging a pre-trained
possessing specialized knowledge and intellectual capac- reinforcement learning algorithm[15], it focuses on iden-
ity collaboratively tackle intricate real-world challenges tifying key players within the co-authorship milieu. By
through self-organizing coordination, thereby propelling combining the complex network topology theory, the study
the advancement of knowledge domains [1]. Within this distinguishes between provincial hub scientists—those cen-
context, the scientific collaboration network constitutes a tral within their respective communities—and connector
fundamental component of the overall innovation frame- hub scientists who bridge different communities. Moreover,
work, embodying the interactive and cooperative dynamics it delves deeper to discern and analyze the multifaceted club-
among researchers [2, 3, 4, 5]. Ongoing development in like properties and functions of members within these two
knowledge engineering and the science of science discipline core structural typologies. The results of this study promise
center on unraveling the nature of collective collaborative to enrich our comprehension of the intricate collaborative
behavior, uncovering emergent collaborative patterns, and patterns in a large-scale social innovation ecosystem.
elucidating the underlying mechanisms driving knowledge
creation system [6, 7, 8, 9].
Existing studies has demonstrated that co-authorship net- 2. Dataset
works typically exhibit typical heterogeneity, confirming
that these networks feature a high degree of uneven distribu- This study focuses on the field of physics. We use the scien-
tion in connectivity [10, 11, 12]. This implies that scientists tific publications in the journals of the American Physical
with extensive social ties wield significant influence over Society (APS) from the period 1985 to 2009 [16]. After the
the network as a whole, often engaging preferentially in col- necessary pre-processing procedure, the dataset finally con-
laborations with other highly influential peers, thus giving tains 104,484 researchers and their 848,231 edges established
rise to the formation of "rich clubs" [13, 14]. by coauthorship relations.
Despite this, the investigation into the diversity of pivotal
actors within expansive scientific collaboration networks 3. Results
remains underexplored, particularly concerning the identi-
fication of mesoscale core structures that bolster global effi- In this study, we propose an interpretable framework to
ciency within large-scale social systems. There is a dearth detect and analyze crucial core structures within large-scale
of research addressing how researchers with varying levels co-authorship networks, integrating previously mentioned
of social capital or differing types of social linkages con- research concepts alongside club structure detection algo-
tribute to the social division of cognitive labor in scientific rithms. This approach involves applying a second-stage
communities. "key player" detection algorithm, which ranks nodes in the
co-authorship network based on their "criticality."
Joint Workshop of the 5th Extraction and Evaluation of Knowledge Entities
from Scientific Documents and the 4th AI + Informetrics (EEKE-AII2024),
As depicted in Figure 1a-c, the network resilience experi-
April 23 24, 2024, Changchun, China and Online ment shows the significance of the key members detected
*
Corresponding author. using the deep reinforcement learning algorithm. Figure 1a
$ liufeifan@dlut.edu.cn (F. Liu); shuang94@mail.dlut.edu.cn illustrates the ratio of the maximum connected subgraph
(S. Zhang); hxxia@dlut.edu.cn (H. Xia) size to the potential maximum size after systematically re-
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribu-
tion 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
106
moving nodes according to their ranked criticality. The
sheer size and complexity of the entire co-authorship net-
work make visualizing it challenging. Therefore, Figures 1b-
c present embeddings of the co-authorship network graphs
between select communities (7, 10, and 14) to exemplify the
influence of key members on the network’s architecture.
The findings reveal that eliminating just the top 28% of key
members causes a near-total collapse of the network, reduc-
ing the maximum connected subgraph size to almost zero.
These results suggest a three-phase impact of key members
on the overall network resilience. From Figures 1b-c, it be-
comes evident that the "key members" recognized by the Figure 1: Structural Characteristics of Connector and Provincial
deep reinforcement learning algorithm play a significantly Hub in Scientific Collaborations Identified by Pre-trained Rein-
more pivotal role in maintaining the network structure com- forcement Learning Algorithm
pared to randomly chosen nodes.
To assess the overlap between the "key members" and the
"pivotal players" in the co-authorship network and verify
of co-authorship networks’ macroscopic and mesoscopic at-
if they support one another, the study conducts statistical
tributes, including scale-free, small-world, modularity, and
analyses. Given that real-world networks tend to display hi-
club structures. While the modularity-based and collective
erarchical modularity [17, 18], we calculate the modularity
collaboration aspects of these networks have received sub-
of each collaborative community, partitioning them further
stantial attention, the in-depth analysis of core structures
into sub-communities using a co-authorship network com-
within co-authorship networks from the modularization and
munity detection algorithm [19]. We then identify "pivotal"
collaboration perspective remains an open issue.
roles within these sub-communities. With an average mod-
Recent research has demonstrated that mesoscopic core
ularity of around 0.71 across the 20 sub-communities, this
structures have indeed been detected and studied in various
suggests a prevalent hierarchical modular organization pat-
domains like biology, transportation, and power systems
tern within the co-authorship network.
[20, 21, 22], playing a pivotal role in global information in-
Moreover, the multi-scale hierarchical modular structure
tegration and subsystem coordination. This study extends
observed in the co-authorship network reflects the inher-
this line of inquiry by exploring the existence of similar
ent hierarchical structure of domain knowledge. Research
mesoscopic core structures in co-authorship networks and
directions, topics, subfields, and disciplines compose the
analyzing their associated network structural traits and func-
knowledge hierarchy in a discipline, and researchers adap-
tional implications.
tively form co-authorships that embed research problems
By harnessing the interpretability of complex topology
within different knowledge system scales.
theory and the representational power of deep learning
Figure 1d presents the variation of club coefficients within
techniques, this study introduces an interpretable frame-
the sub-communities of the multi-scale physical domain as
work to identify and analyze the key cohesive structures
a function of the proportion of deleted nodes (f). It demon-
in co-authorship networks. The study reveals the coexis-
strates that each sub-community contains both global con-
tence of two distinct core structures: local provincial hubs
nector hubs and local provincial hubs, with global connector
that primarily consolidate community members with sparse
hubs exhibiting a stronger cohesive core structure relative
interconnections among themselves, and global connector
to provincial hubs from the complex network system view.
hubs that act as bridges between researchers across different
Figures 1e-f summarize the density and number distribu-
research areas within the collaborative community, main-
tion of "pivotal role" members in the "key member" sequence
taining tight interconnections.
groups. Key observations include: 1) A significant majority
These two types of hubs exhibit minimal overlap and
of globally and locally pivotal members are concentrated
possess unique network structural characteristics, exerting
in the initial sequence subgroups of "key members." This
varying degrees of influence on other network members.
indicates that the higher the criticality rank, the greater
The provincial hubs demonstrate a star-shaped, centralized
the proportion of "pivotal role" members. 2) There is a de-
structure, whereas the connector hubs showcase a flatter
scending order correlation between the criticality of "pivotal
and less centralized pattern of close collaborations.
role" member classification. 3) Nodes with high degree are
The coexistence of local centrality and global decentral-
more critical and occur in larger numbers across both the
ization in co-authorship networks reflects a delicate balance
global collaborative communities and the sub-communities,
between cost-effectiveness, stability, and flexibility within
demonstrating a consistent pattern in terms of importance
the large-scale researcher-driven knowledge exploration
and centrality within the network structure.
process. Future research aims to delve into the potential
universal patterns of scientific meso-core structures across
4. Discussion and Conclusion various disciplines and career stages, drawing upon com-
prehensive academic datasets covering multiple fields and
Scientific collaborative behavior is a cornerstone of large- historical periods.
scale knowledge exploration among researchers and signifi-
cantly influences their academic productivity and impact.
Co-authorship networks serve as a primary analytical tool Acknowledgments
for deciphering collaboration patterns among researchers
This work is supported by the National Natural Science
within a knowledge landscape. As network science theories
Foundation of China (Grant No.71871042 and 72371052).
and methodologies have evolved, so too has the examination
107
References 50. doi:10.1073/pnas.1520118112.
[15] C. Fan, L. Zeng, Y. Sun, Y. Y. Liu, Finding key players in
[1] A. Zeng, Z. Shen, J. Zhou, J. Wu, Y. Fan, Y. Wang, H. E. complex networks through deep reinforcement learn-
Stanley, The science of science: From the perspective ing, Nature Machine Intelligence 2 (2020) 317–324.
of complex systems, Physics Reports 714-715 (2017) doi:10.1038/s42256-020-0177-2.
1–73. doi:10.1016/j.physrep.2017.10.001. [16] S. American Physical, Aps data sets for research, 2018.
[2] M. E. J. Newman, Coauthorship networks and pat- URL: https://journals.aps.org/datasets.
terns of scientific collaboration., Proceedings of the [17] A. Trusina, S. Maslov, P. Minnhagen, K. Sneppen, Hi-
National Academy of Sciences of the United States of erarchy measures in complex networks, Physical Re-
America 101 Suppl (2004) 5200–5. doi:10.1073/pnas. view Letters 92 (2004). doi:10.1103/PhysRevLett.
0307545100. 92.178702.
[3] Z. Xie, Z. Ouyang, J. Li, E. Dong, D. Yi, Mod- [18] P. L. Krapivsky, S. Redner, Emergent network modular-
elling transition phenomena of scientific coauthor- ity, J. Stat. Mech (2017). doi:10.1088/1742-5468/
ship networks, Journal of the Association for Infor- aa7a3f.
mation Science and Technology 69 (2018) 305–317. [19] V. D. Blondel, J. L. Guillaume, R. Lambiotte, E. Lefeb-
doi:10.1002/ASI.23935. vre, Fast unfolding of communities in large net-
[4] S. Yu, F. Alqahtani, A. Tolba, I. Lee, T. Jia, F. Xia, works, Journal of Statistical Mechanics: Theory
Collaborative team recognition: A core plus ex- and Experiment 2008 (2008) P10008–P10008. doi:10.
tension structure, Journal of Informetrics 16 1088/1742-5468/2008/10/P10008.
(2022) 101346. doi:https://doi.org/10.1016/j. [20] R. F. Betzel, J. D. Medaglia, D. S. Bassett, Diver-
joi.2022.101346. sity of meso-scale architecture in human and non-
[5] A. Hajibabaei, A. Schiffauerova, A. Ebadi, Women human connectomes, Nature Communications 9
and key positions in scientific collaboration net- (2018). doi:10.1038/s41467-017-02681-z.
works: analyzing central scientists’profiles in the ar- [21] S. Kojaku, M. Xu, H. Xia, N. Masuda, Multiscale
tificial intelligence ecosystem through a gender lens, core-periphery structure in a global liner shipping
Scientometrics 128 (2023) 1219–1240. doi:10.1007/ network, Scientific Reports 9 (2019) 404–415. doi:10.
s11192-022-04601-5. 1038/s41598-018-35922-2.
[6] Y. Ding, Scientific collaboration and endorsement: [22] C. Seguin, O. Sporns, A. Zalesky, Brain network com-
Network analysis of coauthorship and citation net- munication: concepts, models and applications, Na-
works, Journal of Informetrics 5 (2011) 187–203. ture Reviews Neuroscience 24 (2023). doi:10.1038/
doi:10.1016/j.joi.2010.10.008. s41583-023-00718-5.
[7] W. Li, T. Aste, F. Caccioli, G. Livan, Early coauthorship
with top scientists predicts success in academic careers,
Nature Communications 10 (2019) 5170. doi:10.1038/
s41467-019-13130-4.
[8] Y. Zhu, L. Quan, P.-Y. Chen, M. C. Kim, C. Che, Pre-
dicting coauthorship using bibliographic network em-
bedding, Journal of the Association for Information
Science and Technology (2022). doi:10.1002/asi.
24711.
[9] D. Aguirre-Guerrero, R. Bernal-Jaquez, A methodol-
ogy for the analysis of collaboration networks with
higher-order interactions, Mathematics 11 (2023).
doi:10.3390/math11102265.
[10] A. L. Barabási, H. Jeong, Z. Néda, E. Ravasz, A. Schu-
bert, T. Vicsek, Evolution of the social network
of scientific collaborations, Physica A: Statistical
Mechanics and its Applications 311 (2002) 590–614.
doi:10.1016/S0378-4371(02)00736-7.
[11] R. Guimera, Team assembly mechanisms determine
collaboration network structure and team perfor-
mance, Science 308 (2005) 697–702. doi:10.1126/
science.1106340.
[12] A. M. Petersen, Quantifying the impact of weak,
strong, and super ties in scientific careers, Pro-
ceedings of the National Academy of Sciences of the
United States of America 112 (2015) E4671–E4680.
doi:10.1073/pnas.1501444112.
[13] V. Colizza, A. Flammini, M. A. Serrano, A. Vespig-
nani, Detecting rich-club ordering in complex net-
works, Nature Physics 2 (2006) 110–115. doi:10.1038/
nphys209.
[14] M. Szell, R. Sinatra, Research funding goes to rich
clubs., Proceedings of the National Academy of Sci-
ences of the United States of America 112 (2015) 14749–
108