Graph Neural Network¶
Figures are from: Jia, Z., Lin, S., Gao, M., Zaharia, M., & Aiken, A. (2020). Improving the Accuracy, Scalability, and Performance of Graph Neural Networks with Roc. Proceedings of Machine Learning and Systems, 2, 187–198.
Structure¶
Each GNN layer:
- gathers the activations of the neighbor vertices from the previous GNN layer
- then updates the activations of the vertex, using DNN operations such as convolution or matrix multiplication.
Formally, the computation in a GNN layer is:
\[ a_v^{(k)} = \text{AGGREGATE}^{(k)} (\{h_u^{(k-1)} \mid u \in \mathcal N(v)\}) \]
Note
This corresponds to the left part of the figure.
\[ h_v^{(k)} = \text{UPDATE}^{(k)} (a_v^{(k)}, h_v^{(k-1)}) \]
Note
This step is usually some DNN operations
