Mining The Mode Of Scientific Collaboration Based On Scholarly Big Data

With the rapid development and increasing complexity of science and technology,it is increasingly difficult for scholars to discover new knowledge and propose new theories,which results in researchers’ increasing knowledge and skills to do their research.Collaboration among team members is also becoming an important organization for researchers.And cooperation among team members is conducive to complementing each other’s strengths,promoting the dissemination of research results,and improving the imbalance of scientific development between regions.However,the motivation for collaboration formation,the formula of successful academic collaboration,and the mechanism of scientific team formation are not yet clear.Based on more complete academic data in Computer Science,we construct the collaboration networks by research fields and paper levels,and analyze and model the importance of scientific collaboration,the model of research teams,and the scientific cooperation mechanism in collaboration networks from macro,me-so,and micro perspectives,respectively.The main results and innovations achieved in this thesis are as follows:1)This thesis proposes the academic independence of scholars based on the local topology to discover kernel research teams,which provides a basis for exploring the mechanism of scientific team construction.Based on the characteristics of research teams in organizational behavior and team management theories,the academic independence of scholars based on Ego networks is proposed to determine whether scholars are core members of their scientific teams.And the task of identifying core research teams from scientific collaboration networks is split into two parts: core team member identification and research team detection.When identifying the core members,a handful of scholars published most papers and got numerous attention? this proves the presence of the Pareto principle in academic networks.In the network of the kernel research network which only consists of 30% scholars,most of the ”leaders” and ”members” who belong to these 34 computer vision research teams can be accurately identified,which illustrates that the algorithm can accurately identify the core members of the research teams.In the task of team detection.In the team detection task,a label propagation method based on local similarity is proposed to fully consider the research interests of scholars without setting the size and number of research teams,the members of a group are closely connected.2)A multi-view graph clustering algorithm is proposed,which effectively extracts graph-level structural features and provides a basis for exploring the relationships between research team performance and topology structure.Based on the topology of scientific research teams,a multi-view graph clustering model based on α-Quasi Cliques is proposed based on dense subgraphs of graphs,and graph-level feature extraction is performed on the topology of the same scientific research team under different views to achieve unsupervised graph structure clustering.When clustering the research teams based on the topological structure,the proportions of high-impact and low-impact teams in different categories are different.Some categories show the aggregation of high-impact or low-impact teams,which indicates that team structure affects the performance of research teams.3)This thesis transforms the collaborator recommendation to a link prediction problem in homogeneous networks and proposes a path-based probability estimation model that effectively utilizes the structural information of paths in the network.To address the loss of structural information caused by the cumulative aggregation of information from multiple paths between nodes by existing link prediction algorithms,a path-based probabilistic estimation model is proposed to estimate the potential connectivity between any two nodes.Based on the definition of the probability of connectivity probability,the connectivity of intra-community nodes and inter-community nodes is analyzed qualitatively from the perspective of network communities.The theoretical analysis results show that the predictability of links between intra-community nodes is higher,and the conclusion is validated on artificial data sets(Multi-barbell network)and real network data(Lesmis network).Considering the model’s practicality and efficiency,an iterative algorithm based on adjacency and transfer matrices is applied to simulate paths of different lengths.The experimental results show that the model outperforms the comparison algorithm.4)This thesis proposes a scholar feature embedding model based on dynamic heterogeneous information fusion,which effectively fuses topological structure information and attributes information of entities to predict the value and existence of cooperative relationships simultaneously.By analyzing and organizing the correlations between multiple entities in the scientific collaboration networks,a scholar feature embedding model based on dynamic heterogeneous networks is proposed.It takes into account the topological structure characteristics of the collaboration network while extracting the attributes of heterogeneous entities.In addition,this model is optimized by collaborative optimization to achieve prediction of scientific collaboration potential along with collaboration recommendation.In other words,it simultaneously predicts the value and existence of cooperative relationships.To improve the generalization and practicality of the model,the model is trained with random inputs of the temporal collaboration heterogeneous network constructed by the sliding window method.In the experiment,the results show that fusing different types of entity information can effectively improve the prediction accuracy of the model,and the classification accuracy on the collaborator recommendation task is improved by more than 3%compared with the graph neural network algorithm and the heterogeneous graph neural network algorithm,and the lowest test error is achieved in the collaboration potential prediction task.This thesis recognizes the kernel research teams with a mesoscopic perspective,redefines and identifies core research teams from collaboration networks,explores the relationships between research teams and their topologies;recommends collaborators for scholars with a micro view,and simultaneously predicts the level of papers produced by the collaboration.We explore cooperation patterns in research cooperation networks in a comprehensive perspective,which can be widely applied in other scientific tasks such as paper recommendation,scholar recommendation,academic search,and so on.It will promote cooperation and communication among scholars effectively.