Research On Information Synthesis Based Network Representation Learning Methods

Network is an important form of data used to express the relationship between things and things.With the rapid development of information technology,more and more data appear in the form of networks.Analysis methods such as social networks,word co-occurrence networks,bio-information networks,communication networks,social structure networks,and other large-scale and complicated network structures have attracted increasing attention.Effective network analysis can have a deeper understanding of the knowledge behind the complex network structure,so as to uncover hidden structural information and different correlation patterns between different things,so that it can play a key role in many application scenarios,such as node classification,node clustering,node recommendation,link prediction.The use of machine learning techniques to analyze complex network structure data is plagued by the lack of associated information,sparse data,and high computational costs.For the same task on the same data set,different feature representation methods can result in even larger performance gaps even if the same algorithm is used.Therefore,learning the characteristics of the network structure has become a key research task in network analysis.The network representation learning method expresses the network node information and the correlation information between nodes as a continuous low-dimensional dense vector and then used for the input of various tasks,thereby reducing the computational cost and solving the problem of semantic calculation between heterogeneous objects.However,there are still many challenges in the study of network representation learning.First of all,the working assumptions of most existing networks represent learning in a simple and static environment,but in reality,in most practical problems,the network is not always simple and static.Over time,the network structure will follow the link.The emergence and disappearance of nodes and dynamic changes.The appearance of new content and the disappearance of old content have resulted in changes in the attribute information carried by some nodes.However,existing methods can not process dynamic information structurally,which prompts us to seek an effective knowledge representation to capture the evolution pattern of network topology and node attribute information and eliminate the noise generated by the original model due to the addition of new information.In addition,in most networks representing learning,only the directly connected nodes are considered.For nodes that are not directly connected,there is no way to express the uncertainty of the relationship between them,and it is impossible to effectively capture the complex relationship between nodes.The resulting non-independent related nature.Moreover,the existing network representation model considering additional information generally considers additional information from a single source.For the multi-source and multi-dimensional additional information,the model is difficult to obtain an effective joint vector representation due to the inconsistency of the geometric structure and the acquisition method of the information representation.In order to solve the above problems,this paper has improved the existing network representation learning method based on the existing work.The following are the main work of this paper:(1)A dynamic network representation learning method based on relevance walk and spatial synthesis is proposed.Firstly,parallel embedded space is generated by correlated walk sampling to represent new information and eliminate noise.Then the perturbation factor is considered in the parallel embedded space based on matrix perturbation theory.Integrate into the original semantic space to obtain a low-dimensional vector representation of the updated network.The effectiveness of the method was proved by experiments.(2)A multi-source additional information representation method based on Gaussian distribution is proposed.Firstly,the nodes are embedded with a Gaussian distribution to measure the degree of node association uncertainty.Then a model framework that can embed multi-source additional information in a complementary manner is proposed to embed multiple embedded spaces.Complementarity is embedded into a unified space,and experiments validate the effectiveness of the proposed method.