Link Prediction Of Complex Networks Based On Graph Embedding

Regarding the research of link prediction,there are many models have been proposed.The earliest methods are heuristic algorithms which based on similarity and the latest methods are based on graph embedding.The earliest similarity-based heuristic algorithms are difficult to give a general definition of similarity for different networks,so their performance may be great different in various networks.With the development of graph embedding technology,more and more people apply graph embedding to solve the link prediction problem,but there are still some challenges in the process of using graph embedding for link prediction: during the process of graph embedding,it is unable to treat different order neighbor nodes with different weights when using high-order neighbor features.After getting the representation of nodes,it is unable to take into account the symmetry of edge representation and the integrity of local structure information.This paper focuses on the above problems,and the specific work is as follows:(1)This paper proposes a novel method to aggregate neighbors in different orders with different weights to solve the problem that all orders of neighbors are viewed equally in node embedding.Specifically,the weight matrix is first constructed by aggregating adjacency matrix with different power corresponding to different weight.Then the weight matrix serves as an input for propagating and updating a GNN to obtain the node representations.Due to that the ith row in the weighted matrix contains all the weights of multi-order neighbors for updating the ith node,our method could use min-batch forward propagation to train the GNN while other methods have to use the whole graph data as the input.The experimental results show that using the graph embedding method proposed in this paper can better complete link prediction task by fusing the features of high-order neighbor nodes.(2)This paper proposes a novel method to effectively combine node representations to solve the problem that embedded edges can not keep both local structure information and symmetric presentation simultaneously.Specifically,node pairs are first concatenated both forwardly and backwardly to obtain initial bipartite edges with full local information.Then,these bipartite edges are trained separately through a DNN to obtain their representations.After that,each pair of trained bipartite edge representations are combined through a mix operation to guarantee the symmetry of the final embedded edges.These embedded edge are later used for training the DNN to predict unobserved links.The experimental results show that applying the edge embedding method proposed in this paper can further improve the performance of link prediction,especially in undirected graph link prediction tasks containing symmetric data.