Research on the Evolution and Prediction of the Research Leadership Network in Research Collaborations from the Perspective of Multi-proximities

多元鄰近視角下科研合作領導力網絡的演化機制及預測研究

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
  • Jiang Wu (External person) (External Supervisor)
  • Qingpeng ZHANG (Supervisor)
Award date12 Oct 2021

Abstract

In the “Big Science” era, research collaboration has become the most prevalent and vital form of innovation. Research leadership is essential in the increasingly complex innovation cooperation. The collaboration project is not a simple linear accumulation of tasks but a synergy of various tasks with priorities, logic, and interactions. Therefore, research leadership is indispensable in integrating all parties’ research resources and capabilities and coordinating all parties’ interests and processes. However, the existing studies mainly focus on its qualitative discussion of the types and characteristics, such as the concept, distribution pattern, influence, and other aspects, and less on the quantitative analysis, especially using complex network theory and method. Quantitative interpretation of the research leadership by revealing its network topology features, spatial features, network evolution mechanism, and link prediction can shed some light on the research at the macro national level, the meso institutional/ enterprise level, and the micro individual level.

To this end, this study leverages publication data in pharmaceutical sciences to build the research leadership network at the individual level. From the perspective of multi-proximities, this paper further combines organizational leadership theory, transaction cost theory, resource dependence theory, complex network theory, social capital theory, and homophily theory, and adopts the social network analysis, spatial statistics and visualization, regression analysis, exponential random graph model, network representation learning to interpret the evolution mechanism of the research leadership network and perform link prediction. The main contents are detailed in the following three parts:

(1) The topological and spatial features of the research leadership network. First, we introduce the dataset and preprocessing to construct a research leadership network at the individual level. Second, we analyze the topological and spatial features of the research leadership network. Specifically, for exploring topological features of the research leadership network, we start from two aspects:1) macro network features and 2) micro-network features. The macro network features include the number of nodes, the number of edges, the network density, the average path length, the average clustering coefficient, the network reciprocity, and the network entropy. Micro network features include node degree, weighted out-degree, betweenness centrality, and closeness centrality. In general, the convergence and the polarization of the research leadership network in pharmaceutical sciences are increasing. For analyzing spatial features of the research leadership network, we start from two aspects: 1) research leadership mass, 2) research leadership flow. The spatial analysis on research leadership mass includes spatial aggregation and evolution, and the spatial analysis on research leadership flow includes spatial pattern and evolution. In general, for research leadership mass, the spatial aggregation of research leadership mass is intensifying, with three main clusters.

(2) The role of multi-proximities in edge strength of the research leadership network. In this chapter, we verify the influence of the independent effect and interaction effect of multi-proximities (geographical, cognitive, institutional, and social proximity) on the edge strength of the research leadership network. First, based on resource dependence theory and transaction cost theory, we develope seven hypotheses about the independent and interaction effect of the multi-proximities on the edge strength of the research leadership network. Second, with publications data during 2010-2019 in pharmaceutical sciences, we take the strength of research leadership flow as the dependent variable and geographical, cognitive, institutional, and social proximity as independent variables, and clustering coefficient and research leadership mass as control variables. We further adapt the negative binomial regression model to test the proposed hypotheses empirically. Finally, we conduct robustness tests with fractional counting of the dependent variable and traditional linear regression OLS. For the independent influence of multi-dimensional proximities, geographical proximity has a significant role in promoting the edge strength of the research leadership network. Cognitive proximity has a significant and positive effect. And the cognitive proximity has not formed an inverted “U”-shaped cognitive lock. Institutional proximity has a significant role in promoting the edge strength of the research leadership network. As for social proximity, higher social proximity can promote the edge strength of the research leadership network. As for the interaction effect of geographical proximity and other proximities, higher geographical proximity and cognitive proximity can improve the understanding, processing, and absorption of existing knowledge based on reducing the cost of close interaction and promoting the intensity of interaction. Higher geographical proximity and institutional proximity reduce the uncertainty and enhance the cohesion of behavior. However, the influence of geographical proximity and social proximity is not significant. Once a stable social relationship has been established between researchers, the hindrance of geographical distance will no longer be significant.

(3) The mechanism of the edge formation of research leadership network. In this section, we adopt ERGM to study the formation and evolution mechanism of the research leadership network from the perspectives of network endogenous structural features and node exogenous attribute features. We analyze the following two steps: First, reviewing the relevant literature and proposing research hypotheses. For node exogenous attribute features, we review the related research on the influence of geographic, cognitive, institutional, and social proximity on the edge formation of research collaboration networks. We propose hypothesis H1 of the influence of the multi-proximities on the formation and evolution of the research leadership network. We refine hypothesis H1 into four sub-hypotheses from the four dimensions. Regarding the network endogenous structure features, we review the related research on the preferential attachment mechanism, triadic closure mechanism, reciprocal mechanism on the formation and evolution of complex networks and further proposes the above mechanisms for the formation and evolution of the research leadership network H2-H4. Second, we construct the ERGM model to verify the above hypotheses. We select the publication data in pharmaceutical sciences from 2010 to 2019 as samples to build a directed, unweighted research leadership network. We further explore the influence of multi-proximities, preferential attachment mechanism, triadic closure mechanism, and reciprocal mechanism on the formation and evolution of the constructed research leadership network to verify the research hypotheses.

(4) Link prediction of research leadership network. We address how to comprehensively and systematically integrate node attribute information and network features to conduct research leadership recommendations. A PRLR model is proposed to address these challenges. PRLR integrates cognitive, geographical, and institutional proximity as the node attribute information, capturing the dynamic cognitive base, the geographical proximity, and the institutional proximity of researchers. PRLR constructs a leadership-aware coauthorship network to preserve the research leadership information. PRLR learns the node attribute information, the local network features, the global network features with the autoencoder model, the joint probability constraint, and the attribute-aware skip-gram model. Extensive experimental results demonstrate that the proposed PRLR achieves significant gains over the state-of-the-art collaborator recommendation models. Ablation studies and analysis further show the necessity of integrating geographical and institutional proximities to conduct research leadership recommendations.

    Research areas

  • Research collaboration, research leadership, Social network analysis, Link prediction