Analysis of $p$-Laplacian Regularization in Semi-Supervised Learning