Research On Heterogeneous Network Representation Learning Mechanism Applied For Author’s Academic Behavior Prediction

In today’s academic society,scientific research activities tend to be diversified,multi-cooperation,and interdisciplinary.Academic social network has become an important data resource containing massive academic information.The author’s academic behavior prediction aims to mine the behavioral relationships of authors from heterogeneous academic networks to promote scientific research cooperation and produce high-quality research achievements.Heterogeneous academic networks have the characteristics of high dimensionality and sparseness of data,diversity of nodes and edges,which lead to poor performance of traditional methods.At the same time,most methods do not consider the various features of nodes and the importance of edges between nodes,etc.,which leads to insufficient learning and it is difficult to effectively learn node representations in the network.Therefore,how to effectively fuse multiple features of nodes,mine deep structural features,learn node representations in heterogeneous networks,and efficiently achieve link prediction has become a research hotspot in the field of data mining in heterogeneous networks.In order to effectively extract and fuse various features and neighbor information of nodes,and solve the problems that traditional methods are difficult to deal with heterogeneous network data and do not consider various node features,a heterogeneous network representation learning method based on meta-path and multiple features(HNEMF)is designed.HNEMF considers content features,structural features and community features,and uses bi-directional long short-term memory(BiLSTM)and attention mechanism to fuse the various features and neighbor information of nodes under multiple meta-paths.At the same time,the clustering algorithm is used to capture the global structure,and fully exploit the potential relationship between nodes,so as to effectively learn the representation of nodes and improve the author’s academic behavior prediction accuracy.In order to deeply mine the structural information and semantic information of nodes and solve the problems that most existing methods sample neighbor imbalance,need to pre-define the meta-path to sample neighbor nodes,ignore the importance of edges between nodes and other problems,a heterogeneous network representation learning method based on deep structure(HNEDS)is proposed.HNEDS captures neighbor information through two balanced walk algorithms based on edge information,and uses knowledge graph embedding representation to strengthen the learning of first-order neighbor information.In addition,by mapping nodes to different edge representation spaces for learning,the semantic incompatibility of edges of different types is solved,so that nodes can be represented in depth,the original topology in the network can be preserved to the greatest extent possible,and the effect of author’s academic behavior prediction is improved.In order to effectively extract and fuse node content features,community features and deep structural features,and solve the problems that most of the existing methods only learn node information from a single angle,and do not consider how to effectively fuse multiple features and other problems.On the basis of HNEMF and HNEDS,a heterogeneous network representation learning method based on deep structure and multiple features(HNESF)is implemented.HNESF fuses the content features,community features and deep structural features of nodes through fully connected neural network.Then,the existing knowledge graph embedding technique is improved,the influence of various node features on node representation is increased,the potential relationships among content features,community features,and deep structural features are mined,and the comprehensiveness of first-order neighbor information learning is enhanced.Thus,the learning nodes are comprehensively represented,and the academic behavior prediction performance is further improved.