Research On Graph Embedding Learning Of Multi-features

Graphs(networks)can naturally express objects and their relationship,which has been a common language for modelling complex systems.Graph embedding learning,known as network representation learning,is a process of extracting features of highdimensional system data including structure data and attribute data,and encoding them into low-dimensional dense vectors of real numbers.We can design fast and efficient algorithms based on graph embedding to carry out system data mining.However,with the continuous operation of systems,the scale of system data increases rapidly,which brings considerable challenges to the traditional graph embedding models.In order to reduce computational complexity and learn graph embedding more efficiently,researchers are constantly exploring new methods.Studying graph embedding by deep learning technology has recently become a research hotspot.In this thesis,the graph embedding has been systematically carried out as followings.1.Study on effective attributed graph embedding with information behavior extraction: Most homogeneous graph embedding methods consider structural and attribute features to obtain a node embedding but ignore information behaviour features,including information inquiry,interaction,sharing,and so on.We propose a novel graph embedding model,named the information behaviour joint embedding model(IBE),that incorporates structural features,attribute features,and information behaviour features of nodes into a joint node embedding.At the same time,the impact of the topic number and the feature amplification factor on the performance of IBE is deeply studied.Experiments demonstrate that our model achieves significant improvements in link prediction.2.Research on relation-aware weighted embedding for heterogeneous graphs:Random sampling is used to extract relational semantic features in most existing heterogeneous graph embedding.It loses essential features when the number of random sampling is insufficient,as most graphs do not obey uniform distribution.Although it is possible to get node importance through attention mechanisms developed using expensive recursive message-passing,which can reduce feature loss of important nodes.But GAT adopts a message recursion mechanism,and the features of each node aggregate from a large number of neighbours,which is challenging to deal with large-scale graphs.To resolve this issue,we propose a relation-aware weighted embedding model for heterogeneous graphs(R-WHGE).R-WHGE comprehensively considers structural information,semantic information,meta-paths of nodes and meta-path-based node weights to learn effective node embeddings.Experiments demonstrate that our model has been a significant improvement in node classification.3.Study on weighted sampling enclosing subgraphs embedding for link prediction:Graph embedding methods for large-scale graphs suffer high computation and space costs,and sampling enclosing subgraphs is a practical yet efficient way to obtain the most features at the least cost.Nevertheless,the existing sampling techniques may lose essential features when the random sampling number of nodes is not large,as node features are assumed to follow a uniform distribution.In this thesis,we propose a novel enclosing subgraph embedding model named Weighted Sampling Enclosing-subgraph Embedding(WSEE),which maximumly preserves the structural and attribute features of enclosing subgraphs with less sampling.Experiments demonstrate that WSEE can scale to larger graphs with acceptable overhead while link prediction performance is unaffected.Graph embedding learning aims to map high-dimensional graph data into lowdimensional embeddings and maximize the preservation of the original graph information,which can be used to handle larger datasets effectively.This thesis uses deep learning technology to extract and fuse the graph structure and attribute features,and then encode them into node embeddings.Meanwhile,link prediction and node classification are used to verify the effectiveness of embeddings.