Research And Application Of Zero-Shot Learning Algorithm Based On Knowledge Graph

Currently,classic deep learning tasks such as imageclassification rely heavily on large-scale and strongly annotated data.Zero-shot learning is an effective way to solve this problem,but most existing zero-shot learning methods only use primary knowledge such as attribute information and abstract knowledge such as visual semantic mapping and data distribution,while neglecting the use of knowledge graphs containing a large amount of explicit knowledge.Additionally,the knowledge in the knowledge graph,represented in the form of triplets,suffers from low computational efficiency and sparse data,which is not conducive to downstream tasks.Graph neural networks,on the other hand,are neural network models designed for processing and representing graphstructured data.Furthermore,current research on knowledge graph-based zero-shot learning leans more towards theoretical and academic research,lacking practical applications.In this regard,this thesis proposes the following contributions and innovations:(1)The thesis designs and implements a bidirectional graph convolutional aggregation network based on relationship feature enhancement and Performer.Relationship feature enhancement enhances the embedding representation of knowledge graph nodes,and Performer attention modules reduce the computational time complexity.Moreover,adding the inner-to-outer dot product attention aggregation branch alleviates the problem of internal knowledge dilution.Finally,the model outperforms traditional graph neural networks such as GCN and GAT,as well as Transformer-based graph convolutional aggregation network Tr GCN,on downstream tasks where the learning objective is the weight vector.(2)The thesis designs and implements a knowledge graph-based zero-shot learning method,from subgraph construction and neighbor node sampling to the final classifier learning for unseen classes.The method constructs second-order full subgraphs tailored for different zero-shot learning tasks through class node mapping and differentiated construction processes.The method combines cosine similarity,triple confidence,and random walk sampling mechanisms to maximize the retention of original topological structure information and semantic-related information.Furthermore,adding a two-stage training process for CNN models based on a generated visible-class classifier alleviates the domain shift problem.The final model achieves at least a 1.5% increase in classification accuracy compared to previous models on the AWA2 dataset’s ZSL task and outperforms existing models in the comprehensive indicators for the ZSL and GZSL tasks on the ImageNet dataset.(3)The thesis designs and develops a knowledge graph based zero-shot imageclassification application system,which applies the proposed algorithm model and software engineering techniques to provide users with a knowledge graph based zero-shot imageclassification application system.