Provable Guarantees of Learning with Incomplete and Latent Data