截至这篇 paper 完成之前的大部分基于 message-passing 的 GNN 在表征能力上的上界为 1-Weisfeiler-Lehman (1-WL) graph isomorphism test,这意味着 GNN 无法预测节点集聚系数,最短路径距离和 d -regular graphs。基于此本文提出 Identity-aware Graph Neural Networks (ID-GNNs),核心在于优化 message-passing 的过...
(1)做了一个egonets的操作 defego_nets(graph,radius=2):'''get networks for mini batch node/graph prediction tasks'''# color centeregos=[]n=graph.num_nodes# A proper deepsnap.G should have nodes indexed from 0 to n-1foriinrange(n):ifradius>4:egos.append(graph.G)else:egos.append(nx...
今天学习斯坦福2021年Jure Leskovec发表在AAAI的工作《Identity-aware Graph Neural Networks》。 由Jure Leskovec组 2019 年的工作《How Powerful are Graph Neural Networks?》证明GNN 算法的表达能力,具有与 Weisfeiler-Lehman 图同构测试一样的表达上限,是不适用于计算集聚系数(cluster coefficient),求图内最短距离问题...
Identity-aware graph neural networks. AAAI, 2021. 概 提出了一种能够超越 1-WL-Test 的 GNN. ID-GNN ID-GNN 的 motivation 主要如下: 主要到, 传统的 MPNN, 即第 kk 层: m(k)u=MSG(k)(h(k−1)u),h(k)v=AGG(k)({m(k)u,u∈N(v)},h(k−1)v).mu(k)=MSG(k)(hu(k−1)...
论文链接:FEW-SHOT LEARNING WITH GRAPH NEURAL NETWORKS Idea: Few-shot 学习不依靠于规则化扩展数据集来补偿缺少有监督数据对模型带来的影响,而是受启发与人类学习去探索相似任务在空间上的分布情况。 GNN网络实际是CNN网络在非欧几里得空间上的扩展。 Problem Set-up 以下的定义可以应用于semi-supervised... ...
Specifically, the DiDAL extracts intra-instance salient features by synthetic complementary attention, and further explores the discriminative identity features by modeling the relationship among these salient features based on graph neural networks. This strategy aims to reduce the intra-instance redundancy...
In this paper, the intuition is thattaking the position of nodes in a computational graph to distinguish the node itself. To achieve this, the solution is two-fold: Rather than a layer-wise operation, ID-GNN will first extract the subgraph (ego network around the node) and identity whether...