Implications of sparsity and high triangle density for graph representation learning