Graph Neural Networks
Introduction
我们要怎么把一个graph塞进neural network中呢？之前我们的输入可能是一张照片、一个序列，如果我们把上面这样的图喂给model，要如何让model知道图的结构，也就是node和edge的特征呢？
为什么我们需要用GNN呢？
— GNN可以用在Classification、Generation等问题中。
如何利用卷积将节点嵌入到一个特征空间中？
— Solution 1：将convolution (corelation)的概念推广到图中 → Spatial-based convolution
— Solution 2：回到信号处理中对convolution的定义 → Spectral-based convolution
Roadmap
Tasks, Dataset and Benchmark
Tasks:
- Semi-supervised node classification
- Regression
- Graph classification
- Graph representation learning
- Link prediction
Common dataset:
- CORA: citation network. 2.7k nodes and 5.4k links
- TU-MUTAG: 188 molecules with 18 nodes on average
Spatial-based GNN
Review: Convolution
Spatial-based Convolution
Aggregate：用neighbor feature更新下一层的hidden state
Readout：把所有nodes的feature集合起来代表整个graph
NN4G (Neural Networks for Graph)
DCNN (Diffusion-Convolution Neural Network)
DGC (Diffusion Graph Convolution)
MoNET (Mixture Model Networks)
- 定义节点间"距离"的度量方
- 用加权和（平均值）而不是简单地直接把邻居节点相加（求平均）
GraphSAGE
- Aggregation: mean, max-pooling or LSTM
GAT (Graph Attention Networks)
GIN (Graph Isomorphism Network)
……
