GNN → graph neural network

Introduction

Pytorch Geometric Temporal是Pytorch Geometric的temporal GNN扩展库。

  • 目的

处理Spatio-temporal signals
**

  • 应用

traffic measurements, regional epidemiological reporting or weather

  • 以下temporal GNN模型被Pytorch Geometric Temporal库实现

Discrete **Recurrent Graph Convolutions
节点级预测**

  1.   - **[DCRNN](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.recurrent.dcrnn.DCRNN)** from Li _et al._: [Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting](https://arxiv.org/abs/1707.01926) (ICLR 2018)
  2.   - **[T-GCN](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.recurrent.temporalgcn.TGCN)** from Zhao _et al._: [T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction](https://arxiv.org/abs/1811.05320) (IEEE TITS 2019)
  3.   - **[A3T-GCN](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.recurrent.attentiontemporalgcn.A3TGCN)** from Zhu _et al._: [A3T-GCN: Attention Temporal Graph Convolutional Network for Traffic Forecasting](https://arxiv.org/abs/2006.11583) (CoRR 2020)

链路预测

  1.   - **[GC-LSTM](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.recurrent.gc_lstm.GCLSTM)** from Chen _et al._: [GC-LSTM: Graph Convolution Embedded LSTM for Dynamic Link Prediction](https://arxiv.org/abs/1812.04206) (CoRR 2018)

链路预测、边分类、节点分类

  1.   - **[EvolveGCNH](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.recurrent.evolvegcnh.EvolveGCNH)** from Pareja _et al._: [EvolveGCN: Evolving Graph Convolutional Networks for Dynamic Graphs](https://arxiv.org/abs/1902.10191) (AAAI 2020)
  2.   - **[EvolveGCNO](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.recurrent.evolvegcno.EvolveGCNO)** from Pareja _et al._: [EvolveGCN: Evolving Graph Convolutional Networks for Dynamic Graphs](https://arxiv.org/abs/1902.10191) (AAAI 2020)

场景图分类

  1.   - **[DyGrEncoder](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.recurrent.dygrae.DyGrEncoder)** from Taheri _et al._: [Learning to Represent the Evolution of Dynamic Graphs with Recurrent Models](https://dl.acm.org/doi/10.1145/3308560.3316581) (WWW 2019)

没看

  1.   - **[GConvGRU](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.recurrent.gconv_gru.GConvGRU)** from Seo _et al._: [Structured Sequence Modeling with Graph Convolutional Recurrent Networks](https://arxiv.org/abs/1612.07659) (ICONIP 2018)
  2.   - **[GConvLSTM](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.recurrent.gconv_lstm.GConvLSTM)** from Seo _et al._: [Structured Sequence Modeling with Graph Convolutional Recurrent Networks](https://arxiv.org/abs/1612.07659) (ICONIP 2018)
  3.   - **[LRGCN](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.recurrent.lrgcn.LRGCN)** from Li _et al._: [Predicting Path Failure In Time-Evolving Graphs](https://arxiv.org/abs/1905.03994) (KDD 2019)

Temporal Graph Convolutions
节点级预测

  1.   - **[STGCN](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.convolutional.stgcn.STConv)** from Yu _et al._: [Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting](https://arxiv.org/abs/1709.04875) (IJCAI 2018)
  2.   - **[ASTGCN](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.convolutional.astgcn.ASTGCN)** from Guo _et al._: [Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting](https://ojs.aaai.org/index.php/AAAI/article/view/3881) (AAAI 2019)
  3.   - **[MSTGCN](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.convolutional.mstgcn.MSTGCN)** from Guo _et al._: [Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting](https://ojs.aaai.org/index.php/AAAI/article/view/3881) (AAAI 2019)


Auxiliary Graph Convolutions
节点级预测**

  1.   - **[TemporalConv](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.convolutional.stgcn.TemporalConv)** from Yu _et al._: [Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting](https://arxiv.org/abs/1709.04875) (IJCAI 2018)
  2.   - **[DConv](https://pytorch-geometric-temporal.readthedocs.io/en/latest/modules/root.html#torch_geometric_temporal.nn.recurrent.dcrnn.DConv)** from Li _et al._: [Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting](https://arxiv.org/abs/1707.01926) (ICLR 2018)
  • Recurrent Graph Convolutions概述

    来自《Chickenpox Cases in Hungary: a Benchmark Dataset for Spatiotemporal Signal Processing with Graph Neural Networks》

Recurrent graph convolutional neural networks are highly effective machine learning techniques for Spatio-temporal signal processing.

The specific recurrent graph neural network models fuse ideas from the design of graph convolutional neural network layers and recurrent neural networks.

Recurrent and graph convolutional layers are trained jointly on a downstream task and the design of these architectures requires the choice of a graph neural network and a recurrent unit.
Popular choices for graph neural networks are spectral graph convolutions and graph attention networks while the most frequently augmented recurrent neural networks include long short-term memory cells and gated recurrent units.

RGN融合了GCN和RNN的设计思想。
设计**Recurrent Graph Convolutions架构需要选择GNN和RNN：**

  • 常用GNN
    • spectral graph convolutions
    • graph attention networks
  • 常用RNN

long short-term memory cells（LSTM）
gated recurrent units

DCRNN（ICLR2018 - 节点级预测）

  • 任务

Spatio-temporal forecasting — Traffic forecasting

  • 任务目标

The goal of traffic forecasting is to predict the future traffic speeds of a sensor network given historic traffic speeds and the underlying road networks.
给定历史交通速度和道路网络，预测未来的交通速度

  • 任务建模
    • 将交通的空间相关性建模为有向图上的扩散过程，该有向图的节点是传感器，边权重表示传感器对之间的临近程度。
    • 给定先 前从道路网络上的N个传感器观察到的交通流量。

将传感器网络表示为加权有向图G=(V, β, W)，其中V是结点集合，|V| = N；β是边的集合；
W∈R是表示结点临近度的加权邻接矩阵。
将从图G上观察到的交通流量表示为图信号X∈R，其中P是每个节点的特征数量。
设X代表在时间t观察到的图形信号，交通预测问题旨在学习一个函数h(·)，该函数将T’个历史图信号映射到未来的T个图信号。
image.png

  • 空间相关性建模

使用了本文提出的Diffusion Convolutional Layer，一种GCN网络

  • 时间相关性建模

使用了Gated Recurrent Units（《Empirical evaluation of gated recurrent neural networks on sequence modeling》），一种RNNs

  • DCRNN

With both spatial and temporal modeling, we build a Diffusion Convolutional Recurrent Neural Network (DCRNN).

  • 数据集
    • METR-LA
    • PEMS-BAY

GC-LSTM（CoRR2018 - 动态链路预测）

  • 动态链路预测

动态链路预测，被定义为基于历史网络信息预测未来的网络结构。
Compared with static link prediction, dynamic one is much more difficult since network structure evolves over time.
网络结构会随着时间的推移而演变

  • 应用

在社交网络中，我们预测人们的关系，比如谁将在不久的将来成为谁的朋友。
在通信网络中，我们预测未来的网络结构，
在科学网络中，我们研究研究人员的合作，以预测他们未来的同事。

  • Graph Convolution Embedded LSTM

动态链路预测[7]，[8]被定义为基于历史网络信息预测未来的网络结构。
论文提出的网络，将GCN嵌入到LTSM，实现了端到端的动态链路预测。

GCN in this new deep model is capable of node structure learning of network snapshot for each time slide, while LSTM is responsible for temporal feature learning for network snapshot.
在GC-LSTM中，GCN能够针对每个时间片学习网络快照的节点结构（空间相关性建模），而LSTM负责网络快照的时间特征学习（时间相关性建模）
**

  • 数据集
    • CONTACT and HYPERTEXT09
    • ENRON and RADOSLAW

EvolveGCN（AAAI 2020）

  • 提出先前的GCN+LSTM的局限性，即需要整个时间跨度内的节点知识，并且很难保证将来在新节点上的性能。

In practice, in addition to the likelihood that new nodes may emerge after training, nodes may also frequently appear and disappear, which renders the node embedding approaches questionable, because it is challenging for RNNs to learn these irregular behaviors. To resolve these challenges, we propose instead to use the RNN to regulate the GCN model (i.e., network parameters) at every time step. This approach effectively performs model adaptation, which focuses on the model itself rather than the node embeddings. Hence, change of nodes poses no restriction（节点的更改没有限制）. Further, for future graphs with new nodes without historical information, the evolved GCN is still sensible for them.

  • EvolveGCN

EvolveGCN通过使用RNN evolve GCN的参数来捕捉图序列的动态性。

  • 应用

链路预测、边分类、节点分类

DyGrEncoder（WWW 2019）

Most of these studies have relied on learning to represent nodes and substructures in dynamic graphs. However, the representation learning problem for entire graphs in a dynamic context is yet to be addressed.
这些研究大多依赖于学习表示动态图中的节点和子结构。然而，动态环境下整张图的表示学习问题还有待解决。

We demonstrate that our approach is capable of learning the representation of a dynamic graph through time by applying the embeddings to dynamic graph classification using a real world dataset of animal behaviour.
可以应用到使用“真实世界动物行为数据集”的动态图分类