Research on Knowledge Graph Completion Method Based on Graph Convolutional Neural Networks

Embedding-based knowledge graph complementation techniques have gained more apparent results. However, they still need to improve in dealing with the problem of large-scale sparse graph complementation, and graph convolutional neural network-based knowledge graph complementation methods can better s...

Full description

Saved in:
Bibliographic Details
Published in2024 9th International Conference on Intelligent Computing and Signal Processing (ICSP) pp. 658 - 664
Main Authors Wang, Yunyi, Feng, Shiyu, Wang, Renjie, Feng, Ping
Format Conference Proceeding
LanguageEnglish
Published IEEE 19.04.2024
Subjects
Adaptation models
aggregated neighborhood information
Aggregates
Convolution
Convolutional neural networks
Filters
graph convolutional neural network
knowledge graph complementation
Knowledge graphs
Measurement
Nonhomogeneous media
representation learning
Semantics
Vectors
Online AccessGet full text

Cover

More Information
Summary:Embedding-based knowledge graph complementation techniques have gained more apparent results. However, they still need to improve in dealing with the problem of large-scale sparse graph complementation, and graph convolutional neural network-based knowledge graph complementation methods can better solve this problem and have high research value However, the existing graph convolutional neural network models need to be improved in dealing with node information and global information. Therefore, this paper proposes to dynamically aggregate neighborhood information based on graph convolutional neural networks, where the model uses a single trained weight matrix per layer to aggregate neighborhood information; unlike previous work on convolution using global filters, this paper proposes to adaptively construct convolution filters from relational representations and apply these filters between entity representations to generate convolutional features improve the expressive power of the model; in the datasets, MRR metrics on both FB15k-237 and WN18RR are improved to different degrees compared to the baseline model.
DOI:10.1109/ICSP62122.2024.10743659