Research On Graph Convolutional Neural Network Based Scientific Research Cooperation Prediction

The prediction of scientific research collaboration has always been a prevalent research subject in the field of library and information,along with the scientific research collaboration of all disciplines is becoming more frequent.Scientific research collaboration prediction is one of the tasks of link prediction,which aims to predict the connecting possibility of two nodes that have not yet generated an edge,and further mine the potential collaboration relationship.In this paper,the scientific research cooperation network is made up of nodes and edges,the nodes symbolize authors and the edges symbolize their cooperation relationship.Therefore,these nodes own diverse attributes such as the number of publications,influence index,research topic,gender and so on.Meanwhile the structural features of nodes are contained in the associations among them.In this paper,we analyze the mainstream methods of scientific research cooperation prediction and find they have two shortcomings including:1)the neglect of nodes inherent attribute characteristics leads to low prediction accuracy;2)the failure to achieve "end-to-end" learning results in cumbersome prediction steps.Therefore,in order to make more accurate and efficient scientific research cooperation prediction,this paper applies the rising deep learning method called graph convolutional neural network to predict,which directly acts on network structure to achieve feature learning of nodes.Firstly,a two-layer graph convolutional neural network prediction model is constructed in this paper.The model directly takes the adjacency matrix and feature matrix of one scientific research cooperative network as input,through two times convolution and activation operations,it outputs two nodes similarity score as prediction result.Compared with the method that only makes use of the similarity of network topology structure such as Common Neighbors,Resource Allocation,etc,it additionally integrates the attribute features of nodes to do more accurate prediction.In addition,compared with the network representation learning method such as Deep Walk,LINE,etc,the graph convolutional neural network method learns nodes structure features and attribute features meanwhile.That is to say,the structural and attribute features are put into a network layer for simultaneous learning,and thus the two kinds of features can together influence the final vector representation of nodes.Moreover,the method combines the learning task with the prediction task,which called "end-to-end" learning method.In conclusion,the graph convolutional neural network is an excellent learning method for graph data on account of it can simultaneously learn the attribute and structural features of nodes and achieves "end-to-end" learning.Secondly,the scientific research cooperative networks studied in this paper have heterogeneous edges as a typical weighted network.Therefore,based on the graph convolutional neural network method,we discuss the influence of different weighting ways on the prediction effect.Finally,we selected the Topic-coauthor datasets from the Aminer platform for two contrasting experiments.The first experiment result shows that compared with the traditional link prediction method based on the similarity of network topology,the graph convolutional neural network method has higher prediction accuracy.In the second experiment,by comparing the prediction effects of different linking ways,it can be concluded that different link weights will distinctly lead to different prediction results.So scholars can select appropriate weighting ways to improve the prediction accuracy.