CIKM '21 | PyTorch Geometric Temporal: Spatiotemporal Signal Processing with Neural Machine Learning Models

[jasperzhong]

https://arxiv.org/pdf/2104.07788.pdf

[jasperzhong]

论文基本没讲什么东西..

说了下他们采用的数据结构是discrete temporal snapshot. 节点数量要求不能变. 考虑了三种情况

• dynamic graph with temporal signal: D = {(G1, X1), (G2, X2), ..., (G_T, X_T)}
• dynamic graph with static signal: D = {(G1, X), ... (G_T, X)}
• static graph with temporal signal: D = {(G, X1), ..., (G, X_T)}

看样子他这里的说的signal，应该是node features.

3.2.1说了他们提供了一个Spatiotemporal Signal Iterator的接口. 每次迭代这个iterator返回一个graph snapshot.

具体例子

没太懂为啥用edge_index和edge_attr作为模型输入. 得看代码.

另外paper提到了incremental和Cumulative两种训练方式. 害，前者就是每个snapshot就更新一次模型，后者是每个epoch遍历完了整个dataset再更新一次模型. 后者对显存要求更高，而且从实验效果来看，效果也不好. 如下图中的Wikipedia数据集.

[jasperzhong]

https://pytorch-geometric-temporal.readthedocs.io/en/latest/notes/introduction.html

https://github.com/benedekrozemberczki/pytorch_geometric_temporal

每个snapshot就是一个pytorch_geometric.data.Data，包括

• data.x: Node feature matrix with shape [num_nodes, num_node_features]
• data.edge_index: Graph connectivity in COO format with shape [2, num_edges] and type torch.long
• data.edge_attr: Edge feature matrix with shape [num_edges, num_edge_features]
• data.y: Target to train against (may have arbitrary shape), e.g., node-level targets of shape [num_nodes, ] or graph-level targets of shape [1, ]