Structure-adaptive graph neural network with temporal representation and residual connections
Structure-adaptive graph neural network with temporal representation and residual connections
复制标题
DOI：
10.1007/s11280-023-01179-7
复制
发表时间：
2023-07
期刊：
影响因子：
--
通讯作者：
Xin Bi;Qingling Jiang;Zhixun Liu;Xin Yao;H. Nie;George Y. Yuan;Xiangguo Zhao;Yongjiao Sun
中科院分区：
文献类型：
--
作者：
Xin Bi;Qingling Jiang;Zhixun Liu;Xin Yao;H. Nie;George Y. Yuan;Xiangguo Zhao;Yongjiao Sun
图学习方法促进了对用户医疗保健、疾病检测和行为建模的大脑分析。空间分离的大脑区域以不同的权值进行功能连接，从而可以从图学习的角度对大脑网络进行分类。然而，现有的基于图神经网络的方法主要依靠计算节点特征相关性和人工阈值选择来获得图结构，忽略了节点的时间特征和隐式图结构中的潜在信息。为了解决这一问题，我们提出了一种具有时间表征和残差连接的结构自适应图神经网络（TR-SAGNN）用于脑网络分类。首先，我们设计了一个时间注意学习模块来学习节点本身的时间特征。设计了基于产品矩自注意机制的端到端自适应图结构学习模块，避免了人工阈值选择，获得了更准确的图结构。其次，设计了基于残差连接策略的图表示学习模块，避免了节点特征传播不足的问题；最后，我们设计了一个同时考虑图分类任务和节点分类任务的损失函数，使得模型在节点分类标签的监督下获得了更好的图表示学习能力。我们在ANDI数据集上进行了广泛的实验。结果表明，该模型具有较好的端到端自适应图构建能力以及特征学习和分类性能。
Graph learning methods have boosted brain analysis for user healthcare, disease detection, and behavioral modeling. Spatially separated brain regions are functionally connected with different weights, enabling the classification of brain networks from the perspective of graph learning. However, existing methods based on graph neural networks mainly rely on the calculation of node feature correlation and manual threshold selection to obtain the graph structure, which disregards the temporal features of nodes and the latent information in the implicit graph structure. To address this problem, we propose a structure adaptive graph neural network with temporal representation and residual connections (TR-SAGNN) for brain network classification. First, we design a temporal attention learning module to learn the temporal features of the node itself. We design an end-to-end adaptive graph structure learning module based on the product-moment self-attention mechanism, which avoids manual threshold selection and obtains a more accurate graph structure. Second, we design a graph representation learning module based on a residual connection strategy to avoid the problem of insufficient propagation of node features. Last, we design a loss function to consider both the graph classification task and node classification task, which makes the model obtain better graph representation learning ability under the supervision of the node classification label. We conduct extensive experiments on the ANDI dataset. The results show that our model has better end-to-end adaptive graph construction capability as well as feature learning and classification performance.