Core Concepts
Mini-batch training for graph transformers using personalized PageRank tokenization and virtual connections improves efficiency and scalability.
Abstract
The content discusses the development of VCR-Graphormer, a graph transformer model that utilizes personalized PageRank tokenization and virtual connections to enable mini-batch training. The paper addresses the limitations of dense attention mechanisms in traditional graph transformers and proposes a novel approach to improve efficiency and scalability. Key highlights include:
Introduction of VCR-Graphormer at ICLR 2024.
Challenges with dense attention in graph transformers for large-scale data.
Proposal of personalized PageRank tokenization method for mini-batch training.
Introduction of virtual connections to enhance local, global, long-range interactions, and heterophilous information encoding.
Comparison of complexity between VCR-Graphormer and previous works.
Experiments on small-scale and large-scale datasets showcasing competitive performance.
Stats
VCR-Graphormer needs O(m+klogk) complexity compared to O(n3) in previous works.
Quotes
"Personalized PageRank vector stores the relative importance of other nodes through random walk steps."
"Virtual connections introduce multiple types enabling PPR tokenization to encode comprehensive information."