Patent value prediction is one of the most fundamental components in many applications for knowledge value retrieval. It is important in medical and high-tech fields and crucial in commercial legal proceedings. Since the patent citation relationship has been proven to be one of the most critical factors affecting the value of a patent. Hence, extracting such a citation relationship becomes an essential process. However, extracting the citation relationship between patent documents is still challenging because patent relationship updates frequently. This paper embeds the citation relationship from the patent's Cross Relationship Network (CRN) via the Conditional Variational Autoencoder (CVAE) into prediction tasks. This is the first attempt to leverage the deep generative model to extract a patent's citation relationship into a patent valuation prediction task. Moreover, we also design an interpretable mechanism to judge the goodness. Our approach shows a significant positive effect on refining the prediction of patent value and can enormously assist the Intellectual Property (IP) valuation. We also demonstrate the importance of utilizing patent CRN for extracting the citation relationship. After verifying real-world data, our proposed method shows outstanding results in valuing the newly granted patents by prediction models. In addition, the code and dataset employed in this research will be available to the research community.