Accurate identification of prognostic biomarkers is an important yet challenging goal in bioinformatics. Many bioinformatics approaches have been proposed for this purpose, but there is still room for improvement. In this paper, we propose a novel machine learning-based method for more accurate identification of prognostic biomarker genes and use them for prediction of cancer prognosis. The proposed method specifies the candidate prognostic gene module by graph learning using the generative adversarial networks (GANs) model, and scores genes using a PageRank algorithm. We applied the proposed method to multiple-omics data that included copy number, gene expression, DNA methylation, and somatic mutation data for five cancer types. The proposed method showed better prediction accuracy than did existing methods. We identified many prognostic genes and their roles in their biological pathways. We also showed that the genes identified from different omics data were complementary, which led to improved accuracy in prediction using multi-omics data.