Graph theory has become increasingly significant in machine learning, particularly for tasks involving structured data. This study compares two prominent approaches for node classification–Graph Convolutional Networks (GCNs) as proposed by Kipf and Welling, and Graph Attention Networks (GATs) by Veličković et al.–and generalizes their application to graph classification tasks. Performance evaluation is conducted using sample datasets of varying complexity derived from the Letter Database, which represents handwritten letters as graphs. The findings reveal that GATs outperform GCNs in leveraging adaptive edge weighting, enabling more precise feature integration. Conversely, GCNs exhibit robustness and computational efficiency due to their simpler architecture. This analysis highlights the complementary strengths of these models and their potential in addressing graph classification challenges effectively.