| A network is an abstract representation of real world objects and their interactions.Nodes in a network represent entity objects,and links represent relationships between entities.These links contain rich node attribute information,structure information and network evolution information.Data is missing in the process of network construction.In the process of network evolution,some links may appear or disappear,so it is necessary to complete the missing data and predict the links that may appear or disappear in the future.At the same time,as an important branch of data mining field,link prediction has been applied to social network,e-commerce,science and other fields,which has very important practical significance.At present,most of the similarity based link prediction methods are based on the assumption of node representation,which makes use of the feature of common neighbors of nodes.As a result,the node representation obtained by a single index is relatively similar,and it is difficult to effectively distinguish the node characteristics in the network.At the same time,the pooling of traditional deep neural network makes it difficult to retain the network structure information effectively in the node feature representation,which affects the prediction effect of the model.the main research work of this paper is as follows:(1)For networks without node attributes,a link prediction model based on cascading generalization is proposed.Firstly,from the perspective of machine learning,the similarity index is taken as the feature between node pairs to transform link prediction into a dichotomous problem.Then,by using ensemble learning,the overlapping generalization mechanism is introduced,and different machine learning models are used to effectively combine different similarity indexes,so as to obtain the fusion indexes that better describe the structure characteristics of the target network,thus improving the prediction performance of the model.(2)For the network with node attributes,a link prediction model based on graph neural network and capsule network is proposed.Firstly,the graph neural network is used to learn and generate corresponding node features,and the network nodes are mapped to low dimensional node feature vectors.Secondly,through the designed conversion block,the learned node features are converted into the corresponding node pair features.Thirdly,inspired by the idea of capsule network,the structure information between node pairs can be captured effectively by learning the feature representation between node pairs through capsule network.At the same time,different graph neural networks can be used in the model to adapt to different target networks,so as to improve the prediction performance.(3)A network relationship prediction system for scientific collaborators is designed and implemented.The link prediction model proposed in Chapter 3 and Chapter 4 is applied to the expert cooperation network prediction system,aiming to provide help and suggestions for researchers to get familiar with the network relationship information of their own field and make effective decisions.According to the expert information input by users,the cooperative network relationship is predicted by the model,and the predicted cooperative network relationship results are presented to users in a visual way.To sum up,this paper proposes t link prediction models based on cascade generalization and a link prediction model combining graph neural network and capsule network respectively for the network without node attributes and the network with node attributes.Experimental verification is carried out on many different real data sets,and the existing methods are compared and analyzed.Experimental results show that the proposed link prediction model achieves good prediction performance on several evaluation indexes such as AUC and F1-Score,which verifies the validity of the model.Finally,the model is applied to expert cooperation network system,and good prediction results are also obtained. |
PDF Full Text Request