Research Of Network Representation Learning Method Based On Information Networks

The network representation learning algorithm represents the information network as a low-dimensional dense real vector that carries the characteristic information of network nodes,and is applied to the input of downstream machine learning tasks.With the rapid development of machine learning and deep learning in recent years,network representation learning has more powerful modeling and feature extraction capabilities and has been widely used.The paper has conducted in-depth research and exploration in three aspects,namely classification and review of network representation learning algorithm,improvement of existing algorithms,and classification of the efficacy of network representation learning on TCM prescription data.The current complex network representation learning methods are classified into network node embedding method and graph neural network method,which are subdivided into six categories,namely,traditional method,network structure-based embedding,embedding integrating attribute information,and spectral-based Graph Convolutional Networks,spatial-based Graph Convolutional Networks and Graph Attention Networks.The applicability and characteristics of the network representation learning model are compared,and the open source representation learning model and graph deep learning library are listed,as well as the commonly used data sets for the reference of researchers,so as to provide a comprehensive review reference for scholars in the field of network representation learning.The traditional network node embedding method ignores the original information and attribute information existing in the network,and the graph convolution algorithm has no tradeoff in feature extraction of first-order neighborhood of nodes.Aiming at the above two problems,the Enhanced network representation with weighted edge and feature fusion(FWENR)is proposed.In this method,embedded vector based on network structure are fused with the initial feature vector of the network,so that the initial features of the model contain more structure and attribute information,and the weight calculation mechanism is introduced to the unweighted network,so that the model has a tradeoff in the feature extraction of the node's first-order neighborhood.The data of TCM prescriptions were collected and processed,and the TCM prescription network was constructed based on the co-occurrence relationship of drugs in the prescription,and the initial feature matrix of the network was designed.The experimental results prove that,compared with the structure-based network representation learning and graph neural network methods,the FWENR model has achieved a certain performance improvement in the feature representation and efficacy classification tasks of the prescriptions network.