Multi-mode Data Fusion Based Network Representation Learning And Application

Network can express the relationship between different things,and it is ubiquitous in everyday life,such as social network,transportation networks,the Internet networks and so on.Network representation learning represents each node in the network as a low-dimensional dense vector,which is easy to calculate and can be flexibly used for various network analysis tasks.However,network data is often complex.For example,in social networks,nodes contain user attributes,texts,tags,etc.,and edges contain user interaction information,and different user interactions constitute multiple structures.The above information in different forms constitutes the network multi-mode data studied in this paper.Multi-mode data fusion based representation learning can utilize data association to mine hidden rules and improve the quality of network representation.In view of the current fusion representation learning methods,which don't fully integrate more useful multi-mode data,we propose two multi-mode data fusion based network representation learning methods in this paper.The main contributions are as follows:1)Aiming at the situation that text and multi-structure are mixed in social network,a network representation learning method is proposed,which integrates friend relationship structure,interaction relationship structure and node text.Firstly,as for TADW,which is a network representation method based on matrix decomposition and combining text and structure,this paper analyzed and discussed the influence of the location of text attributes matrix on network representation.Secondly,on this basis,we analyze the relationship between multi-mode data,map the multi-structure and text of the network into a low-dimensional feature matrix,and place them in the optimal position of the decomposition formula,and learn the network representation through matrix decomposition.Finally,the user interest discovery experiment was conducted on the Microblog dataset,and the results show that this method is superior to other classical methods.2)In view of the coexistence of node information,edge semantic information and network structure in the network,we propose a network representation learning method based on automatic encoder and translation mechanism.The model uses multiple automatic coders to represent the information of node and edge respectively,and uses translation mechanism to model nodes and edges to maintain the network structure.The model updates the representation through collaborative optimization,and achieves the effective fusion of multi-mode data.Finally,two kinds of multi-relationship extraction experiments were carried out on the data set collected by the ArnetMiner platform: cooperative multi-relationship completion,and multi-relationship prediction among new authors.Compared with the classical method,the accuracy of this model is improved in two different tasks.It is shown that thisalgorithm not only has a good effect on discovering implicit multiple relationships,but also can predict the multiple relationships introduced by new nodes in dynamic networks without retraining the model.