Batching for GNN

Batching for GNN

• GIven I training graphs $X_i,A_i$ and labels $y_i$, params $\Phi =\{{\beta }_k,{\Omega }_k{\}}_{k=0}^K$ , SGDand Binary Cross Entropy
• We cannot concat batches into a big tensor since each graph has a different number of nodes
• Instead, we treat the graphs in each batch as disjoint components of a single large graph.
• we can then just run the network over this big graph as a single instance of the network equations.
• Then we can do mean pooling on it

Neighborhood Sampling

• randomly sample fixed number of neighbors, recursively
• somewhat like Dropout

Graph Partitioning

• cluster the original graph into disjoint (not connected to each other) subsets of nodes
• these are batches
• this converts a transductive problem to an inductive one
• for inference, k-hop neighbors