Graph data has been integrated into transfer learning-based cross-domain rotating machinery diagnosis for reducing domain discrepancy. Sample relationships, representing the correlations between data distribution and the sample label, have been destroyed in the graph construction process, resulting in transferable knowledge loss. To fully mine and retain the transferable knowledge, a transferable graph features-driven cross-domain rotating machinery fault diagnosis approach is proposed. An improved graph construction strategy is designed to establish the mapping between labels and nodes. Graphs with similar structure for source- and target-domain samples are constructed to preserve sample relationships under data distribution discrepancy. Domain adaptation is introduced to the graph convolutional network for reducing learned graph feature discrepancy. Case studies, including two cross-load and one cross-machine transfer diagnosis tasks, are conducted for effectiveness verification. Experimental results show that it can effectively learn transferable graph features to eliminate the cross-domain discrepancy.