Towards Interpretable Sparse Graph Representation Learning with Laplacian Pooling