| With the explosive development of complex network science,researchers have established complex network models for various real-life systems,and studied their behaviors and characteristics.These behaviors and characteristics cannot be separated from the behavior of individuals in complex systems and the relationships between individuals.Link prediction provides an effective prediction mechanism for mining the interactions and correlations between individuals.A great deal of research has been conducted in the field of link prediction.However,most of the existing methods generally suffer from some problems,for example,many algorithms rely on low-order interaction relationships and are unable to capture the interactions between two nodes in terms of high-order relationships and high-order structures,and many algorithms are limited to predicting pairwise interactions between nodes,and can’t predict high-order relationships between multiple nodes beyond pairwise interactions directly.Therefore,to address the above existing problems and limitations,we take the research on high-order link prediction in ordinary graphs as the foundation,the research on link prediction based on high-order structure as the bridge and transition,and the research on hyperlink prediction algorithm based on high-order structure-hypergraph as the research focus,and carry out the researches on link prediction based on highorder interaction relationship recognition for ordinary graphs and hyperlink prediction for hypergraphs.The main research contents are as follows:(1)Considering that most of the current link prediction algorithms only study the connection between nodes,without considering the interaction between nodes and the high-order relationship between nodes,a novel algorithm framework based on gravitational field and node contraction is proposed from the perspective of physics,and is applied to improve the link prediction performance of the existing link prediction algorithms.Experimental results demonstrate that leveraging high-order relationships between nodes can effectively enhance link prediction performance.(2)Most of the link prediction algorithms proposed so far study the binary interaction relationship between nodes based on the chain structure and star structure,and do not consider the interactions between nodes from the high-order structural information that exists in the network.The cycle structure is a higher-order structure that lies between the star and clique structures,which can describe the higher-order interactions between multiple nodes.Therefore,a multi-information fusion node importance algorithm based on the cycle structure information is proposed,which integrates both high-order and low-order structural information,and then a graph neural link predictor with multi-information interactions based on the cycle structure is developed.Through extensive experimental validation and analysis,the proposed neural network model can effectively learn the cycle structure information,and using higher-order structural information-cycle information proves to significantly enhance both the accuracy of node importance estimation and the overall performance of link prediction.(3)Hyperlink prediction is much more challenging than other downstream tasks in complex networks,as it requires predicting the possibility of multiple nodes(unrestricted)forming hyperedges.Therefore,two preferential hypergraph weighted random walk with restart strategies are proposed and applied to hyperlink prediction.Finally,through extensive experimental verification and analysis,it is demonstrated that the proposed methods have good hyperlink prediction performance on both uniform and non-uniform hypergraph datasets.(4)Most of the existing hyperlink prediction methods predominantly focus on predicting undirected hyperlinks,and how to predict the directed hyperlinks is a difficult and worth studying problem.Based on the above problems,the graph neural network is utilized to learn graph structure data,and a graph neural network-based hyperlink prediction framework is proposed to predict the interaction between multiple targets,which can be applied not only to predict hyperlink prediction in undirected hypergraphs but also in directed hypergraphs.Experimental results show that the hyperlinks can be predicted efficiently by the proposed hyperlink prediction framework with relatively low time complexity. |
