.png){kind=icon}
.png){kind=icon}
返回
Frequency inception based graph neural network for relation prediction in knowledge graphs
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2023.110908.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Recently, knowledge graphs have been broadly studied using approaches such as translation-based models and convolutional neural networks. Although such approaches can express powerful and rich embeddings, graphs are naturally suitable for irregular data. Thus, they can model knowledge graphs with structured data and aggregate considerable information. However, graphs are subject to incompleteness, leading to unconnected knowledge graphs. To address this problem, we propose a framework called frequency-inception-based graph neural network (FiGNN) for relation prediction in knowledge graphs. It exploits a graph to aggregate the most beneficial information through mathematical analysis. Specifically, combined with the relations and inception parameters, each channel of a node and its neighbours dynamically contribute to next-layer channel information, which can extract underlying signals of different channels and neighbouring nodes. Moreover, the representation ability of the nodes is enhanced and over-smoothing is alleviated. We validate the effectiveness of the proposed network on various benchmark datasets. The experimental results demonstrate that our network achieves substantial improvements and substantially outperforms state-of-the-art methods. & COPY; 2023 Elsevier B.V. All rights reserved.
Keyword:
Knowledge graph competition prediction
Deep representation learning
Frequency inception