Understanding Negative Sampling in Graph Representation Learning