Research On Link Prediction Algorithm Of Complex Network Based On Line Embedding Representation

In real life,complex and changeable network systems can be represented as graph data.In recent years,Graph Embedding technology has gradually become one of the important methods for studying complex network graphs,which has important research significance and extensive application value.The Graph Embedding technology is to efficiently extract the feature information of the graph and express it as a low-dimensional dense vector,which can be used in network analysis such as Node Classification and Link Prediction.Link Prediction is to predict potential or undiscovered links in complex networks,which is widely used.Based on the research of graph data,this paper propose two improved Link Prediction algorithms and an improved Graph Embedding algorithm,and conducts comparative experiments on multiple real data sets.The specific research work is as follows:(1)This paper aims at the problem that the transition probability of the basic random walk similarity index was determined just by the degree of current node,which makes the link prediction effect unsatisfactory.In view of this,an improved link prediction algorithm IMRWR(Improved MH with RWR)is proposed.It is based on the MetropolisHasting(MH)algorithm,which comprehensively use the degree information of the current node and neighbor nodes to redefine the transition probability between nodes,and the probability matrix is reconstructed by assigning weighted the self-loop rate of current node to its neighbors.Then the Random Walk with Restart(RWR)similarity index is fused in the proposed method.The link prediction experimental result on a variety of real network data sets show that the consequence of our method IMRWR is better than other seven benchmark algorithms,the AUC index of proposed algorithm also is improved,as well as the ranking score index.(2)Existing Word2 Vec model-based Graph Embedding algorithms fail to invoke the initiative of the sampled nodes in the sampling process,so that the quality of the obtained node sequence was not high,which can influence node embedding effect.This paper proposes a novel graph embedding algorithm Line2 Vec.It can comprehensively consider the joint influence of the current node and neighbor node on the sampling probability and uses a flexible unbiased and biased random walk neighborhood sampling strategy.This way can fully invoke the initiative of the next node to be sampled on the sampling process to enhance the quality of node sequence generated by walking and obtain node vectors with stronger performance.In this paper,Node Classification and Link Prediction experiments show that our algorithm Line2 Vec has higher Micro-F1,Macro-F1 and AUC values,which enhances the representation of node vectors.(3)The node graph was used by most graph embedding algorithms by default,which can only obtain the vector representation of the node and cannot directly obtain the vector representation of the edge.In this paper,the node graph is converted into a line graph and sent the line graph into the newly proposed algorithm Line2 Vec that can process the node graph and the line graph at the same time.The vector representation of the edge can be directly obtained and the vector representation of the non-existing edge can be obtained by combining the incidence matrix used for the first time.In this way,a Line-embedding link prediction algorithm Line2Vec-L is proposed.The directly obtained edge vectors contains the first-hand information of the graph without being diluted or weakened,and is more representative.The Link prediction experiments show that the AUC value of our algorithm Line2Vec-L is higher,and the generated edge vector is more expressive.