Automatic modulation classification (AMC) is a promising technology to realize intelligent wireless communications in the sixth generation (6G) wireless communication networks. Recently, many data-and-knowledge dual-driven AMC schemes have achieved high accuracy. However, most of these schemes focus on generating additional prior knowledge or features of blind signals, which consumes longer computation time and ignores the interpretability of the model learning process. To solve these problems, we propose a novel knowledge graph (KG) driven AMC (KGAMC) scheme by training the networks under the guidance of domain knowledge. A modulation knowledge graph (MKG) with the knowledge of modulation technical characteristics and application scenarios is constructed and a relation-graph convolution network (RGCN) is designed to extract knowledge of the MKG. This knowledge is utilized to facilitate the signal features separation of the data-oriented model by implementing a specialized feature aggregation method. Simulation results demonstrate that KGAMC achieves superior classification performance compared to other benchmark schemes, especially in the low signal-to-noise ratio (SNR) range. Furthermore, the signal features of the high-order modulation are more discriminative, thus reducing the confusion between similar signals.