A Hybrid VAE Based Network Embedding Method for Biomedical Relation Mining

Abstract

Mining biomedical entity association and extracting the implicit knowledge from biomedical entity relation networks are important for precision medicine. In this paper, we propose a novel method for implicit relation mining from biomedical multi-entity network. In the embedding part, we combine two kinds of model (1) the graph representation learning model like GraphGAN and (2) the network embedding model like VAE based SDNE, to construct a hybrid model GVS. In the prediction part, the positive samples selected from original network and the negative samples generated by ranking meta-paths are used to train kNN. To evaluate the performances of GVS, we compare the proposed method with three state-of-the-art methods (Katz, Catapult and IMC) on benchmark datasets. Moreover, we evaluate GVS on a real biomedical entity relation network, it shows advantages compared with other network embedding methods and successfully mines implicit relationships which validated by PubMed.