OpenGT
收藏资源简介:
OpenGT是一个用于图Transformer的综合性基准,它为Graph Transformers提供了统一的实验设置,并包含了多种最先进的GNN和GT模型。OpenGT通过评估GTs在多个角度的性能,包括不同的任务类型、图结构、数据集规模、注意力机制和图特定信息的集成策略,来促进公平的比较和多维分析。OpenGT通过广泛的实验揭示了图Transformer的几个关键洞察,包括跨任务级别转移模型的难度、局部注意力的局限性、几种模型的效率权衡、特定位置编码的应用场景以及一些位置编码的预处理开销。OpenGT旨在为未来的图Transformer研究建立公平性、可重复性和泛化性的基础。
OpenGT is a comprehensive benchmark for Graph Transformers, which provides unified experimental configurations and encompasses a wide array of state-of-the-art GNN and GT models. OpenGT facilitates fair comparisons and multi-dimensional analyses by evaluating the performance of GTs across multiple dimensions, including diverse task types, graph structures, dataset scales, attention mechanisms, and integration strategies for graph-specific information. Through extensive experiments, OpenGT has uncovered several key insights into Graph Transformers, including the difficulty of transferring models across task levels, the limitations of local attention, efficiency trade-offs of several models, applicable scenarios for specific positional encodings, and the preprocessing overhead associated with certain positional encodings. OpenGT aims to establish a solid foundation for fairness, reproducibility, and generalizability in future Graph Transformer research.




