Generalizable Insights for Graph Transformers in Theory and Practice
This work addresses the gap between theory and practice in Graph Transformers for researchers and practitioners, providing generalizable insights, though it is incremental as it builds on existing advancements.
The paper tackled the lack of generalizable insights for Graph Transformers by proposing the Generalized-Distance Transformer and conducting extensive experiments, identifying design choices that perform well across diverse applications and demonstrating strong performance in few-shot transfer without fine-tuning.
Graph Transformers (GTs) have shown strong empirical performance, yet current architectures vary widely in their use of attention mechanisms, positional embeddings (PEs), and expressivity. Existing expressivity results are often tied to specific design choices and lack comprehensive empirical validation on large-scale data. This leaves a gap between theory and practice, preventing generalizable insights that exceed particular application domains. Here, we propose the Generalized-Distance Transformer (GDT), a GT architecture using standard attention that incorporates many advancements for GTs from recent years, and develop a fine-grained understanding of the GDT's representation power in terms of attention and PEs. Through extensive experiments, we identify design choices that consistently perform well across various applications, tasks, and model scales, demonstrating strong performance in a few-shot transfer setting without fine-tuning. Our evaluation covers over eight million graphs with roughly 270M tokens across diverse domains, including image-based object detection, molecular property prediction, code summarization, and out-of-distribution algorithmic reasoning. We distill our theoretical and practical findings into several generalizable insights about effective GT design, training, and inference.