Research On Higher-order Link Prediction Methods Based On Network Representation Learning

With the rapid development of network science and big data technology,complex networks have become powerful theoretical and practical tools for solving problems in various fields.As one of the hot topics in complex networks,link prediction refers to the use of observed network data to predict missing links and future potential links.It is of great significance for understanding the evolution mechanism of networks and mining key factors that affect the dynamic changes of networks.In addition,link prediction has a wide range of practical values in various fields.It is used in protein-protein networks to predict future interactions between proteins.In social networks(such as Twitter and Facebook),it can be used to explore potential friend relationships and make recommendations to increase user stickiness.Furthermore,link prediction can be used in knowledge graph completion and e-commerce networks to recommend products to customers.With the need to model relationships between three or more entities,high-order networks have wider applications in various fields.Interpersonal relationships,such as multi-author cooperation,co-occurrence of keywords,and joint purchasing,can be naturally modeled as high-order networks.However,due to computational complexity and the scarcity of high-order data in datasets,the exploration of high-order networks is usually limited to triads(or triangles).To address these issues,this paper explores and quantifies the similarity between different nodes and edges in a network.The goal of this paper is to solve the high-order problem through biased random walks and use representation learning methods,and to explore the possibility of using non-negative matrix factorization to process the higher-order structure formed by network nodes and edges.(1)A higher-order link prediction method based on the higher-order structure of the nodes is proposed.This method uses the higher-order structure of the nodes around the nodes to the link prediction,so that the feature vector distance can be used to quantify the similarity between nodes and edges in the network.The algorithm uses biased random walk to abstract the higher-order structure information around the nodes and uses the representation learning methods to map the walking sequence of the nodes into the low dimensional feature vectors,so that it can contain the higher-order structural features around the nodes.The method performs well on the prediction of higher order structure in eight real networks than other baseline algorithms.(2)A higher-order link prediction method based on the factorization of the edge relation matrix is proposed.The currently used higher-order link prediction method based on local topology structure can not evade the influence of the network’s disconnectedness.Therefore,we combine non-negative matrix factorization and higher-order link prediction to obtain the similarity between the nodes and edges in the network and to consider the possibility of the relationship between the nodes and edges generating higher-order structure of the network.Compared with traditional algorithms,this algorithm shows excellent performance in real networks.