Automated map generalization at regional scales remains a bottleneck in national mapping agencies. Generalization rules for hydrographic networks are complex and difficult to quantify using direct indicators. To address generalization challenges in drainage and examine how non-visual information can improve feature selection, this study evaluates and adapts an automated Graph Convolutional Network (GCN) approach, focusing on the GraphSAGE model proposed by (Wang and Qian, 2023). The hydrographic network is represented as a graph, with nodes corresponding to drainage segments and edges representing their connections. Segment classification, as retained or discarded, is based on semantic, geometric, morphological, topological, and constraint-related attributes. The method was applied to drainage generalization using four attribute sets derived from the Brazilian Technical Specifications of the Geospatial Vector Data Structure (ET-EDGV). Three geometric attributes, length, sinuosity, and polygon containment, were tested together with one non-geometric attribute: fluvial regime. Using only geometric attributes, the best model achieved 94.52% training accuracy, 94.70% validation accuracy, 93.91% Recall, and a 95.30% F1 Score. When the fluvial regime was included, performance increased to 99.98% test accuracy, 99.98% validation accuracy, 99.98% Recall, and a 99.98% F1 Score. The generalized network was validated against a 1:100,000 reference drainage dataset manually generalized from the same 1:25,000 input, yielding 67.16% positional accuracy for the model trained only with geometric attributes. Results indicate that GraphSAGE is suitable for selecting drainage segments from ET-EDGV attributes, especially for scale transitions from 1:25,000 to 1:100,000.