Overcoming the curse of dimensionality with Laplacian regularization in semi-supervised learning