However, most existing autoencoder-based methods discard the reconstruction of auxiliary information, which poses a huge challenge for better representation learning and model scalability.