| Scholarly impact assessment plays an important role in the process of the dissemination and development of academic research.The impact factor based on the citation distribution calculation is an important indicator for evaluating the influence of entities in heterogeneous scholarly networks.However,the value of the impact factor is jointly affected by sociological and statistical factors.The existing related research is mainly based on the statistical method of the database to compare and analyze the literature network information and data,which can not address and explain the impact of social behavior differences on data distribution.Furthermore,the variation of impact factor is also influenced by the link relationship,interaction behavior and network node information between different entity layers in a heterogeneous scholarly network.Based on the node distribution law of citation network and the calculation mechanism of impact factor,this dissertation studies the interaction between different entities in a heterogeneous scholarly network and impact factor.The main contributions of this dissertation can be summarized as follows:(1)An experimental approach of literature evaluation indicators based on unsupervised Laplacian Score is proposed to analyze the differences of evaluation indicators in discipline categories and quartiles.The Laplacian score algorithm can find the features that reflect the local manifold structure of the dataset.To exploit local and global structural information of the dataset simultaneously,the local and global discriminant criteria LGD based on unsupervised learning and the Fisher Score based on supervised learning are employed to verify the feature selection results.According to the results of feature selection and the available label information,the performance of features with different Laplacian scores in classification and clustering is validated by utilizing spectral clustering and k-NN.Based on the BPNN model,a literature classifier that simultaneously predict discipline categories and quartiles is developed.The experimental results demonstrate that the impact factor with the highest Laplacian score and LGD score achieved the best performance in the process of classification and clustering,which can better reflect the local and global manifold structure features of the dataset.Compared with other evaluation indexes,impact factor is an evaluation index with the most degree of discipline differentiation and quartile differentiation.(2)A weighted ranking evaluation scheme based on impact factor and citation relevance is proposed to improve the link relationships between different sub-networks in heterogeneous scholarly network.In addition,a time-aware method FutureRank is utilized as the feature vector of PageRank algorithm to balance the bias to earlier node in the dynamic citation network.Two ranking evaluation methods(P-Rank and W-Rank)with different iterative execution are selected as the optimization goal of the weighted scheme.The weighted ranking evaluation method here conducted a comprehensive evaluation of node time information,writing,citation and publication among three entity units(author,paper and j ournal)in the heterogeneous scholarly network.The experiments are conducted on three datasets under different network configuration conditions and weighting schemes.The experimental results show that the two weighted algorithms introduced impact factor,citation relevance and time information achieved superior Spearman’s ranking correlation and robustness on the three datasets.(3)In order to analyze the correlation between the interaction behavior of entities in heterogeneous scholarly networks and the variation of impact factor,this dissertation develops a virtual science network community composed of three different entity units(author,paper and journal)based on the parallel system.The submission behavior,peer review behavior,citation behavior and time information of nodes in a heterogeneous academic network are analyzed by employing related functions.According to the calculation mechanism of impact factor and citation distribution,parallel system is utilized to qualitatively analyze the influence of the behavior and decision-making of entity units in different heterogeneous scholarly networks on the distribution of nodes in the citation network and the variation of impact factor.The experimental results demonstrate that the variation of impact factor,the difference of citation behavior and the change of node time information in citation network can be well recognized by the developed parallel system.Furthermore,the parallel system can be used to generate the corresponding virtual citation networks,and the virtual citation networks here are well compatible with the most typical scale-free networks. |
