Link Prediction Based On Graph Attention Network And Its Application Research

Link prediction,a research hotspot in the field of network analysis,aims to use the information known about the network to infer potential or unknown links in the system.Link prediction can solve the problem of incomplete network connectivity,obtain missing data,identify spurious interactions,evaluate network evolution mechanisms,etc.Knowledge graph is a structured representation of facts,consisting of entities and relationships.Using knowledge graphs can integrate scattered information materials and build a powerful semantic information network,which in turn creates good conditions for solving practical problems.Along with the enrichment of knowledge,knowledge graphs need to be updated and improved continuously,and it is always a concern of researchers for how to infer the implicit relationships and potential connections among entities.In this paper,we address the problem of link prediction in various models such as static knowledge graphs,temporal knowledge graphs and multilayer networks,and our main work is as follows:(1)To address the problem that entities are treated in isolation in static link prediction,an extended attention network of relationship graph is proposed to fully capture the features of entity neighbors and mine the hidden information between triads.Meanwhile,a novel embedding method of entities and relations is designed to realize the fusion of entities and relations from multiple perspectives under vector space,which takes into account the potential semantic information of entities and relations at the same time and can realize the fusion embedding of entities and relations.(2)To address the problem of fine-grained fusion of time-frequency features in temporal link prediction,an iterative bilinear fusion method is designed to improve the cross-domain feature representation of entities and relations in a new iterative way.Meanwhile,a temporal convolutional network is introduced to capture the sequential patterns between adjacent facts across time to solve the oversmoothing problem of entity embedding convergence to the same value.In addition,cosine similarity is introduced to measure the similarity between entity features.This work uses a parallel approach to acquire time-frequency features,which effectively realizes the cross-domain interaction between time-frequency features.(3)To address the problem of discrete and isolated processing of "node pairs" in multilayer network link prediction,we propose a two-level graph attention mechanism,which achieves fine extraction of small-scale node neighborhood features by the node-level attention mechanism and fusion of large-scale "node-pair" topological features by the "node-pair" level attention mechanism.The fusion of topological features of a large range of "node pairs" is achieved by the node-level attention mechanism.At the same time,a multi-metric fusion approach is used to more comprehensively characterize the similarity between nodes in the same layer and measure the inter-layer similarity based on the cosine theorem to achieve an effective measurement of the prediction weights of the target layer and multiple auxiliary layers.In addition,an end-to-end multilayer network link prediction framework is designed to integrate "node-pair" feature learning and predictor to effectively improve link prediction accuracy.