Research On Link Prediction Algorithms Based On Topology Structure And Network Representation Learning

In recent years,with the vigorous development of Internet technology,the complex network science that studies the connection between things has also developed rapidly.Link prediction is one of the popular research directions in complex network science.Its main task is to predict missing or unknown edges in the network,which is of great significance in theoretical research and practical application.At present,most link prediction algorithms use the local topology information of the network when calculating the similarity between candidate nodes;the node representation vector obtained through the network representation learning algorithm contains potential network structure features;the random walk algorithm performs link prediction Medium can play a very good effect.Inspired by the above,this paper takes the undirected unauthorized network as the research object,and improves the link prediction algorithm from three angles,aiming to improve the accuracy of the prediction algorithm.Starting from the local topology of the network,on the basis of the CRA index and the CCN index,the local community structure of CRA is extended to the form of common neighbors and three-hop neighbors,and parameters are added to balance the contribution of the two nodes,so that The local community structure depicts the similarity between nodes and proposes a new link prediction index—LXR index.Experiments on 10 real data sets prove that the improved LXR indicator has a good performance.Then consider the similarity relationship between the nodes in the network representation learning algorithm-DeepWalk algorithm node sequence,change the sampling method of the algorithm walk sequence,and change the random sampling to biased sampling.The similarity between the nodes is calculated by the LXR index Characterization,so that the similarity between the nodes in the improved node sequence is higher,so that the resulting low-dimensional vectors of nodes can better characterize the network structure.The improved algorithm is called-G-DeepWalk algorithm.Performing link prediction experiments on 13 real data sets verifies that the algorithm has a good prediction effect;the angle cosine distance can describe the similarity between nodes and vectors more than the Euclidean distance,and it also improves the prediction accuracy.It is more stable and has good robustness;using network representation learning to perform link prediction can improve prediction accuracy better than direct link prediction.Finally,the node representation vector obtained by DeepWalk and the algorithm G-DeepWalk algorithm is first combined,and then the representation vector and the common neighbor similarity are combined and normalized,so that the improved transition probability matrix is applied to restart random walk and local random walk.Algorithm.Experimental results on 20 real data sets show that the improved algorithm has a very good performance in the results of link prediction and has good applicability.