The map generalization process is a complex method aimed at producing small-scale maps from large-scale maps, while preserving the structures and characteristics of geographic features and transforming them into a legible representation at the target scale. River networks, due to their branched, hierarchical, and topology-dependent structure, are considered among the most challenging features in the generalization process. Although data-driven and deep learning methods have been increasingly applied to river network generalization in recent years, many of them do not explicitly model the hierarchical structure of the network or, by disregarding cartographic constraints, lead to inappropriate removal or retention of branches. In this study, a two-stage framework for river network generalization and simplification is proposed. In the first stage, the hierarchical structure of the network is extracted by combining Horton–Strahler ordering with geometric stroke connectivity analysis. In the second stage, the selection and elimination of branches are performed using a graph-based deep learning model with a GraphSAGE architecture, implemented in accordance with the guidelines of the National Cartographic Organization. The proposed model was applied to the river network data of Isfahan Province at a scale of 1:25,000, and the results were compared with the 1:50,000 reference map. To evaluate accuracy, the Coefficient of Linear Conformity (CLC) was used, which measures the longitudinal overlap between the generalized network and the reference map. The results showed that the CLC value was 0.96, indicating appropriate preservation of the overall network structure and a high level of agreement between the model output and manual cartographic generalization at the target scale.