A graph transformer with optimized attention scores for node classification

The message-passing paradigm on graphs has significant advantages in modeling local structures, but still faces challenges in capturing global information and complex relationships. Although the Transformer architecture has become a mainstream approach for nonlocal modeling in many domains, Transformer-based architectures fail to demonstrate competitiveness in popular node-level prediction tasks when compared to mainstream graph neural network (GNN) variants. This can be attributed to the fact that existing research has largely focused on more efficient strategies to approximate the Vanilla Transformer, thereby overlooking its potential in node embedding representation learning. This paper introduces a novel graph Transformer model with optimized attention scores, named OGFormer, to address this gap. OGFormer employs a simplified single-head self-attention mechanism, incorporating several critical structural innovations in its self-attention layers to more effectively capture global dependencies within graphs. Specifically, we propose an end-to-end attention score optimization loss function, designed to suppress noisy connections and enhance the connection weights between similar nodes. Additionally, we introduce an effective structural encoding strategy integrated into the attention score computation, enabling the model to focus more on key local dependencies in the graph. Experimental results demonstrate that OGFormer achieves strong competitive performance in node classification tasks across multiple benchmark datasets, particularly in Homophilous and Heterophilous tests, surpassing current mainstream methods. Our implementation is available at https://github.com/LittleBlackBearLiXin/OGFormer .