Graph Transformers with Attention over Temporal Edges

This paper introduces a novel approach to Graph Transformer architectures by integrating attention mechanisms specifically designed to capture temporal dependencies within graph data. Traditional Graph Transformers have largely focused on modeling spatial relationships between nodes, often overlooking the crucial evolution of these relationships over time. We propose a method to augment the standard Graph Transformer structure with attention layers that explicitly consider temporal edges – edges representing connections that change over time – and their dynamic behavior. This allows the model to learn how relationships shift and adapt, leading to improved performance in tasks requiring understanding of temporal graph dynamics. The core contribution lies in the explicit modeling of temporal edges through an attention mechanism, enabling the network to learn patterns in how connections form and dissolve. This framework provides a more comprehensive representation of graph data compared to traditional spatial-only approaches.