The paper introduces Graph Attention Networks (GATs), a novel neural network architecture designed to process graph-structured data. GATs leverage masked self-attentional layers to address the limitations of previous methods based on graph convolutions or their approximations. By stacking layers where nodes can attend over their neighborhoods' features, GATs enable implicit specification of different weights for different nodes in a neighborhood without requiring costly matrix operations or prior knowledge of the graph structure. This approach addresses key challenges of spectral-based graph neural networks and makes the model applicable to both inductive and transductive problems. The authors evaluate GAT models on four established benchmarks: Cora, Citeseer, Pubmed citation networks, and a protein-protein interaction dataset, achieving or matching state-of-the-art results. The paper also discusses the theoretical and practical benefits and limitations of the GAT architecture compared to prior work, and highlights potential future directions, such as improving batch handling and enhancing model interpretability.The paper introduces Graph Attention Networks (GATs), a novel neural network architecture designed to process graph-structured data. GATs leverage masked self-attentional layers to address the limitations of previous methods based on graph convolutions or their approximations. By stacking layers where nodes can attend over their neighborhoods' features, GATs enable implicit specification of different weights for different nodes in a neighborhood without requiring costly matrix operations or prior knowledge of the graph structure. This approach addresses key challenges of spectral-based graph neural networks and makes the model applicable to both inductive and transductive problems. The authors evaluate GAT models on four established benchmarks: Cora, Citeseer, Pubmed citation networks, and a protein-protein interaction dataset, achieving or matching state-of-the-art results. The paper also discusses the theoretical and practical benefits and limitations of the GAT architecture compared to prior work, and highlights potential future directions, such as improving batch handling and enhancing model interpretability.