Research On Representation Learning Of Complex Network Structure

There are a lot of network structures in real life,such as social networks and ci-tation networks.With the advancement of digitization,the effective use of network structure information has become an important part of data mining.Network represen-tation is a powerful tool to solve this problem.By learning the nodes in the network as a low-dimensional vector representation to benefit downstream tasks,it is currently a research hotspot in data mining.The simplest network data is the homogeneous network data,that is,the nodes be-long to the same category and only have simple structure information,and at the same time,there is only one type of network structure.However,the network data in real data is complex and changeable,and usually exists in the form of a complex network structure.For example,nodes belong to multiple types of information,and the edges between nodes have multiple types.Both the knowledge graph and the attribute net-work belong to a complex network structure.The modeling of the knowledge graph network is usually generated indirectly by the link prediction task of the fact triple,which contains many types of edges and the types of nodes are diverse? and the mod-eling of attribute network representation requires the fusion of attribute information on the basis of a homogeneous network,that is,nodes have both heterogeneous in-formation of structural information and attribute information.This paper studies the representation learning of complex network structures—knowledge graph network and attribute network,and explores the problems in the current work.Based on the basic idea of deep information interaction,it proposes a new solution for the modeling of the above network.The main work of this article is as follows:·Explore the problem of insufficient utilization of prior knowledge in the embedding of knowledge graphs and propose solutions.The knowledge graph is composed of triples,and the relationship among them has a guiding role in the prediction and representation of entities,while previous work has not explicitly modeled this.We propose a method based on the relationship-limiting gate.The model constrains the correct entity to be close to the extra golden representation of the relationship,while the wrong entity far from it.This method explicitly utilizes the constraint ability of the relationship,can expand to multiple models and improves the link prediction task on multiple datasets.·Explore the problem of noise filtering in knowledge graph embedding and propose solutions.The link prediction process of knowledge graph is an inference process in itself.We propose a knowledge graph representation learning algorithm based on the multi-step iterative gate mechanism by simulating the predictive inference process of humans.This method uses the gate mechanism to control the interac-tion between the relationship and the entity,filter useless or noisy information,and strengthen the filtering effect through a multi-step iteration mechanism to obtain better prediction results.This method achieves the optimal effect of link prediction on multiple datasets,and the effectiveness of each module in the model is verified by the ablation study.·Explore the problem of outliers in the attribute network and propose solutions.There are a large number of abnormal nodes in the attribute network,which means the nodes may not belong to the same category as the neighbor nodes with similar structures or attributes.However,many current models are based on this kind of similarity assumption?.In addition,there is inconsistency between the structure and attribute information which means they describe a node differently,and it is urgent to dig deep complementary information to make up for the lack of original information.In this paper,a deep attribute network representation learning model based on neighbor enhancement is proposed.The model efficiently mines effective neighbor information by truncated neighbor sampling and similarity limitation,and proposes a reverse attention mechanism to merge neighbor information effectively.Autoencoders are used to output high-quality representations.The model has excel-lent performance in node classification,node clustering and visualization tasks on multiple datasets.Module analysis and parameter sensitivity analysis further verify the effectiveness of the model.