Open YijianLiu opened 1 year ago
For link prediction, batching is not performed over individual edge buckets, but rather over the in-memory subgraph, which is a union of the edge buckets currently in memory. From the in-memory subgraph, a set of training edges are selected and mini-batches are created over this selection. The reasoning of using this selection process and mini-batch generation are located in the final paragraph of Section 5.1. in our GNN paper.
Batches can be processed synchronously, or asynchronously with some configurable staleness bound.
Let me know if that answers your question.
Whether each bucket will perform mini-batch training? On your paper, whether each bucket performs 4(bound) mini-batch training? Is my understanding correct? Thanks a lot!