![](https://img-blog.csdnimg.cn/20201014180756919.png?x-oss-process=image/resize,m_fixed,h_64,w_64)
GNN
明天又是好吃的一天v
现在在吉林大学读研。
展开
-
[GNN论文阅读]Towards Deeper Graph Neural Networks with Differentiable Group Normalization
NeurIPS 2020Towards Deeper Graph Neural Networks with Differentiable Group Normalization一、主要工作二、研究动机三、创新点四、具体思路(i)两个指标(ii)DGN五、理论分析六、实验效果一、主要工作(本文针对节点分类问题)(i)给出了衡量过平滑的两个指标:1.Gins:衡量输入的embedding特征表示X有多大程度服务输出H。Gins的公式可类比于KL(P(X,H)||P(X)*P(H))相对熵(其实基本原创 2020-12-06 20:01:27 · 1181 阅读 · 0 评论 -
【GNN论文阅读】Residual or Gate? Towards Deeper Graph Neural Networks for Inductive Graph Representation L
Residual or Gate? Towards Deeper Graph NeuralNetworks for Inductive Graph Representation Learning 一、主要工作二、研究动机三、创新点四、具体思路五、理论分析六、实验效果Networks for Inductive Graph Representation Learning )一、主要工作文章提出了一种用于inductive图表示学习的通用gra原创 2020-11-28 20:12:22 · 776 阅读 · 0 评论 -
[GNN论文阅读]DROPEDGE: TOWARDS DEEP GRAPH CONVOLUTIONAL NETWORKS ON NODE CLASSIFICATION
DROPEDGE: TOWARDS DEEP GRAPH CONVOLUTIONAL NETWORKS ON NODE CLASSIFICATIONICLR 2020代码地址: 1.Absrtact1.Absrtact原创 2020-11-20 19:50:03 · 768 阅读 · 0 评论 -
[GNN论文阅读]DeepGCNs: Can GCNs Go as Deep as CNNs?
一、主要工作基本思路和解决的任务:CNN可以深卷而获得较好的性能,然不能表示非欧氏空间的数据。GCN可以表非欧氏空间的图,然因梯度消失(过平滑,节点特征趋于一致)不能深卷(彼时<=4)。此文提出一种多达56层的深卷GCN。利用残差网络ResNet、密集连接DenseNet、空洞卷积We adapt concepts that were successful in training deep CNNs, mainly residual connections, dense connectio原创 2020-11-13 16:07:18 · 415 阅读 · 0 评论