Learning probabilistic latent process models from temporal data