GAT-Edge: Graph Attention Neural Network with Adjacent Edge Features
Abstract: Edge features are a crucial component of graph data as they provide a wealth of information that can enhance model performance. In this paper, we propose an improved model called GAT-Edge, which builds upon the graph attention network by optimizing the attention mechanism to incorporate edge feature information. By leveraging adjacent edge features in the graph, our model can assist downstream tasks such as node classification. The connection between nodes in graph data is often enriched by the adjacent edges, which provide more effective and abundant information. To exploit this, our model combines edge features and node features in the attention calculation, convolving them together to generate new attention coefficients. This approach facilitates efficient information transmission and aggregation between nodes, leading to improved performance. We apply our new model to several citation networks commonly used in the field of graph neural networks for node classification, and compare it with the current mainstream graph convolution neural network models. Our results demonstrate that our model achieved better accuracy, highlighting the importance and research value of mining adjacent edge features in graphs.
Paper Type: Long
Research Area: Machine Learning for NLP
Research Area Keywords: Graph Neural Network, Attention Mechanism, Edge Feature, Node Classification
Contribution Types: Approaches to low-resource settings, Data analysis
Languages Studied: English
Submission Number: 1481
Loading