Research On Link Prediction In Complex Networks With Multi-feature Fusion

As an important research topic in complex networks,link prediction not only has important research significance in the fields of knowledge graph completion and exploring the evolution law of networks,but also has extensive practical value in the fields of friend recommendation,movie recommendation and metabolic network reconstruction.The task of link prediction is to complete the missing link or predict the new link in the future according to the information of the current network structure.Existing link prediction algorithms are often based on network topology modeling,which is highly hypothetical and cannot automatically learn graph structure features,and has low accuracy.In view of the main problems existing in link prediction in complex networks,this paper improves and constructs relevant algorithms from the perspectives of network topology,machine learning,graph representation learning,etc.,so as to comprehensively and multi-level mining of link information in the network,so as to accurately improve the performance of the link prediction algorithm.The main work and innovations of this paper are as follows:(1)Aiming at the problem of low accuracy and strong hypothesis of heuristic link prediction algorithm,a link prediction algorithm based on XGBoost was proposed.In addition to selecting the classical similarity index as the feature of the sample,four new features are defined,namely,the sum of first-order neighbors,the sum of second-order neighbors,the shortest path between nodes,and the local path information of nodes.The above features are input into XGBoost algorithm,and XGBoost algorithm automatically selects the features that adapt to the current network structure.It solves the problem of low precision and strong hypothesis of heuristic link prediction algorithm.Experimental results show that compared with heuristic link prediction algorithm,our algorithm can improve the result of link prediction.(2)In order to solve the problem that the machine learning-based link prediction algorithm needs to extract graph structure features manually and the model cannot learn graph structure features automatically,a link prediction algorithm based on subgraph was proposed.The algorithm analyzes the subgraph of the target node pair and establishes the graph marking algorithm module.The graph marking algorithm module assigns a label to each node in the subgraph,and the full connection layer module can automatically learn the graph structure features by the label.In addition,the algorithm integrates the attributes of nodes,tags of nodes and implicit features of nodes to improve the performance of link prediction algorithm.Experimental results show that the link prediction algorithm based on subgraph is superior to benchmark methods.(3)Different nodes in the network have different importance.Based on this characteristic,a link prediction algorithm based on graph attention network is proposed.The algorithm establishes an attention mechanism to distinguish the importance of different neighbor nodes.Through the convolution layer module,the characteristics of neighbor nodes and discarded nodes are aggregated to make the nodes have richer information representation.The pool layer module defines the relative order of nodes in the subgraph,obtains the expression of subgraph features,and improves the performance of link prediction algorithm.The final experimental results show that the link prediction method based on graph attention network obtains better prediction results than the benchmark method.