The Graph Attention Network (GAT) is a widely recognized architecture in the field of Graph Neural Networks (GNNs). It is considered the state-of-the-art approach for graph representation learning. In recent years, several researchers have successfully applied GAT to structured Euclidean data, including images and languages. Additionally, the Graph Convolutional Network (GCN) has also been adopted in information systems, such as PN-GCN, which establishes undirected graphs for classification. However, compared to undirected graphs, directed graphs contain directional information that more comprehensively describes the relations between objects in the graph. Therefore, we establish a directed graph in the information system and further analyze it using GAT. As a first step, this article introduces the concept of multi-granularity object directed weighted graphs. Then, we construct the second generation of the residual edge-weighted graph attention neural network model (PD-GATv2) based on these directed graphs. Finally, we verify the effectiveness and generalizability of the PD-GATv2 algorithm through experiments, and its effectiveness is further demonstrated through ablation experiments.

图注意力网络（GAT）是图神经网络（GNN）领域广泛认可的架构。它被认为是最先进的图表示学习方法。近年来，一些研究人员已成功将 GAT 应用于结构化欧几里得数据，包括图像和语言。此外，图卷积网络（GCN）也被应用于信息系统中，例如PN-GCN，它建立无向图进行分类。然而，与无向图相比，有向图包含有方向的信息，可以更全面地描述图中对象之间的关系。因此，我们在信息系统中建立一个有向图，并利用GAT对其进行进一步分析。作为第一步，本文介绍了多粒度对象定向加权图的概念。然后，我们基于这些有向图构建第二代残差边缘加权图注意神经网络模型（PD-GATv2）。最后，我们通过实验验证了PD-GATv2算法的有效性和泛化性，并通过消融实验进一步证明了其有效性。