Reseach And Implementation Of Multi-view Network Representation Learning

In recent years,with the emergence of a large number of structured data in the digital age,how to calculate,analyze and mine this kind of more complex information has become an urgent problem.As a method of modeling structured data,information network has been widely used in various scenarios and tasks.Network representation learning,by designing the similarity measure between nodes,can effectively represent the nodes or edges in the network as dense vectors in low dimensional space,and learn rich information such as network topology and attribute characteristics,which solves the problems of high spatial-temporal complexity and poor generalization performance faced by traditional adjacency matrix or feature construction;at the same time,it can also benefit the future using various advanced algorithms and models in machine learning.The learned representation is used as the input feature vector,which can be efficiently applied to various downstream tasks such as relational reasoning,recommendation system,classification and clustering.Therefore,network representation learning can effectively calculate,analyze and mine structured data.Although the existing methods have done a lot of research on the representation learning of homogeneous network(containing only one node type)and heterogeneous network(containing multiple node types)respectively,and have been successfully applied in various tasks,most of the existing methods only focus on solving the single view network,that is,there is only one relationship between a pair of nodes,and can not be effectively extended to other multi-view network.But in reality,the existence of multi-view network is more extensive,such as friend relationship and forwarding relationship in social network,co-authorship relationship and citation relationship in academic network.Therefore,the effect of existing methods in real scene multi-view network still needs to be improved.The information contained in different views formed by different relationship types may be consistent or complementary.Therefore,how to identify and preserve the complementary information in the multi-view network,and how to fuse the consistent information in the multi-view network is the key,so that the learned representation vector can capture more comprehensive,rich and complex multi-view network information.In this paper,the related work of network representation learning based on multi-view is carried out,which can be divided into the following three points:1.A multi-view based network representation learning model MV-ACM is proposed.The network topology similarity calculation module is used to measure the consistency and complementarity of each view.Then,it uses the adversarial learning to generate the representation vector of multi-view network:discriminator module,distinguishes the complementarity among the views,and uses the hierarchical hidden state to extract the information,so as to retain the unique complementary information of each view;generator module,recommends multiple views with stronger consistency,through the shared intermediate hidden state and convolution network,and achieve consistent information fusion.We use the learned representation vector for link prediction and node classification tasks.The experimental results on a variety of real data sets show that the effectiveness of our proposed model is significantly improved.2.A multi-view network representation learning model based on motif is proposed.For single view networks,we find different motif substructures and establish corresponding motif relationships for the nodes in the substructures,so as to form multi-view networks based on different motifs.The original single view network is transformed into a multi-view network with more abundant high-order structure information,and then the multi-view model is used for representation learning.Experiments show it effectively learns the substructure information of the network,and enhances the expression ability of representations.3.We apply the multi-view based network representation learning to the specific task of product search and recommendation on Taobao platform.Through the construction of the double view network of search click and recommendation click between users and products,the search and recommendation data can be unified modeling.At the same time,multi-view network representation learning based on attention mechanism is used to get the representation of users and products,which is used to predict the final click through rate.Through the joint learning of user preferences and product features under the two views,the effect of our model has been significantly improved on both tasks.