Research On Link Prediction Method Based On Graph Attention Network

Knowledge graphs,with powerful knowledge representation capabilities,are widely applied in various fields such as information retrieval,natural language understanding,and recommendation systems.However,the incompleteness of knowledge graphs can result in incomplete search results,inadequate accuracy,and limited application domains,making link prediction an important method for expanding and improving knowledge graphs.In recent research,graph attention network models have achieved good results in link prediction due to their ability to capture local and global information effectively.However,there are still issues in the initialization phase,such as the lack of semantic information for entities and relations,as well as problems in the decoding phase,such as the neglect of triplet structural features and the inclusion of low-quality negative samples.This paper addresses these issues and presents the following main contributions:(1)To address the problem of missing semantic information for entities and relations in the initialization phase,we propose a method that utilizes pre-trained language models to encode entity and relation description information.This method ensures the uniqueness of the embedding representations for entities and relations during the initialization phase,and extracts semantic information from their description,enriching the embedding representations of entities and relations.To address the problem of neglecting triplet structural features in the decoding phase,we propose a decoder based on CNN that concatenates the embedding representations of triplets into matrices and applies convolution and pooling operations,preserving the structural features and important embedding information of the triplets.In the encoding phase,we propose an encoder based on multi-head attention mechanism that aggregates the embedding representations of adjacent entities and relations with weights,enabling the target node to obtain more important information.(2)To address the problem of low-quality negative samples in the negative sample set,we propose two optimization methods for negative sampling.One is an adversarial learning-based negative sampling optimization method that uses a discriminator and a generator to obtain high-quality negative sample sets.The other is a cache-based negative sampling optimization method that designs a cache mechanism for learning from simple samples to complex samples,and samples and updates the cache to obtain a high-quality negative sample set.(3)Design and implementation of a knowledge graph link prediction system.Through research on link prediction methods,we have designed and developed a knowledge graph link prediction system,which has been validated using the Chinese Frame Net(CFN)dataset.In addition to basic functions such as data import,export,visualization,manipulation of nodes and relation,the system applies link prediction methods to the CFN framework relation prediction task,enabling the discovery of potential framework relations.The implementation of this system facilitates the management and maintenance of the CFN dataset and provides technical support for improving the CFN knowledge base.