In the original GraphSAGE implemetation, author does negative sampling for a batch of edges (positive samples). I see sampling in the tutorial but don’t see a reference of negative sampling. Just wondering if there is a provision for that in DGL.
GraphSAGE sampling tutorial is dealing with node classification rather than link prediction.
Currently, DGL is only supporting sampling batches of nodes natively. We aim to provide a more intuitive and flexible sampling framework in 0.5, and we will be supporting more sampling strategies natively by then.
To sample by edges with the current version, currently this PR which runs GraphSAGE on MovieLens for recommender systems temporarily works around the issue by manually preparing the seed node arrays, including negative sampling.