Pytorch Geometric笔记

链接：Fast Graph Representation Learning with PyTorch Geometric

https://arxiv.org/pdf/1903.02428.pdf

定义

图：

X：n个节点，

边特征：

边indices：

Neighborhood Aggregation(message passing)

图卷积操作：

图卷积

⬚ : 不同的function，e.g., sum, mean or max

γ : differentiable function, which can be achieved by gathering and scattering of node features and vectorized element-wise computation
3 params are chosen by users.
**已经配置模型：GCN, SGC, GraphSAGE, the attention-based operators GAT and AGNN, the Graph Isomorphism Network(GIN), the Dynamic Neighborhood Aggregation(DNA) operator.
Also provide maximizing mutual information, auto encoding graphs, aggregating jumping knowledge and predict temporal event s in knowledge graphs.

Pooling

• Global pooling: Graph-level outputs by functions like global add, mean or max pooling
• Hierarchical pooling

Mini-batch Handling

Batches of multiple graph instances (of potentially different size) by automatically creating a single block-diagonal adjacent matrix and concatenating feature matrices in the node dimension.

Application

Semi-supervised classification
Graph classification
Point Cloud classification