基于优化问题下的图神经网络研究
Research on Graph Neural Networks Based on Optimization Problems
文诗琦 1皮家甜2
作者信息
- 1. 重庆师范大学数学科学学院,重庆 401331
- 2. 重庆国家应用数学中心(重庆师范大学),重庆 401331
- 折叠
摘要
为了探究图神经网络各个模型之间的本质关系,更好地运用图神经网络解决实际问题,从基于空域的三个图神经网络出发,在统一的优化框架下,由不同的卷积核得到三个具体的优化问题,求解这些问题得到信息传播机制的表达式.结果表明,各模型的传播机制,分别对应特定图正则化项下优化问题的解,区别在于卷积核的设计不同.本研究为设计新的图神经网络变体提供了思路.
Abstract
In order to explore the essential relationship between various models of graph neural networks and better apply graph neural networks to solve practical problems,starting from three graph neural networks based on spatial domain,under a unified optimization framework,three specific optimization problems are obtained from different convolutional kernels,and the expression of information propagation mechanism is obtained by solving these problems.The results indicate that the propagation mechanisms of each model correspond to the solutions of optimization problems under specific graph regularization terms,and the difference lies in the design of convolution kernels.This study provides ideas for designing new variants of graph neural networks.
关键词
图神经网络/可解释性/传播机制Key words
graph neural networks/interpretability/propagation mechanism引用本文复制引用
出版年
2024
NETL
NSTL
万方数据