Identifying important nodes with strong propagation capabilities in road networks is a significant topic in the field of urban planning. However, existing methods for evaluating nodes importance consider only topological information and traffic volumes, ignoring the diversity of characteristics in road networks, such as the number of lanes and average speed of road segments, limiting their performance. To address this issue, this paper proposes a graph learning-based node ranking method (MGL2Rank) that integrates the rich characteristics of the road network. In this method, we first develop a sampling algorithm (MGWalk) that utilizes multi-graph fusion to establish association between road segments based on their attributes. Then, an embedding module is proposed to learn latent representation for each road segment. Finally, the obtained node representation is used to learn importance ranking of road segments. We conduct simulation experiments on the regional road network of Shenyang city and demonstrate the effectiveness of our proposed method. The data and source code of MGL2Rank are available at https://github.com/ZJ726.