Enhancing Graph Transformers with Hierarchical Distance Structural Encoding

Authors: Yuankai Luo, Hongkang Li, Lei Shi, Xiao-Ming Wu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we demonstrate that graph transformers with HDSE excel in graph classification, regression on 7 graph-level datasets, and node classification on 11 large-scale graphs.
Researcher Affiliation Academia Yuankai Luo1,3 Hongkang Li2 Lei Shi1 Xiao-Ming Wu3 1Beihang University 2Rensselaer Polytechnic Institute 3The Hong Kong Polytechnic University luoyk@buaa.edu.cn xiao-ming.wu@polyu.edu.hk
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Our implementation is available at https://github.com/LUOyk1999/HDSE.
Open Datasets Yes Datasets. We consider various graph learning tasks from popular benchmarks as detailed below and in Appendix B. Graph-level Tasks. For graph classification and regression, we evaluate our method on five datasets from Benchmarking GNNs [20]: ZINC, MNIST, CIFAR10, PATTERN, and CLUSTER. We also test on two peptide graph benchmarks from the Long-Range Graph Benchmark [23]: Peptides-func and Peptides-struct... Node Classification on Large-scale Graphs. We consider node classification over the citation graphs Cora, Cite Seer and Pub Med [44], the Actor co-occurrence graph [14], and the Squirrel and Chameleon page-page networks from [71], all of which have 1K-20K nodes. Further, we extend our evaluation to larger datasets from the Open Graph Benchmark (OGB) [35]: ogbn-arxiv, arxiv-year, ogbn-papers100M, ogbn-proteins and ogbn-products...
Dataset Splits Yes For each dataset, we follow the standard train/validation/test splits and evaluation metrics in [68]. For more comprehensive details, readers are encouraged to refer to [68]. Cora, Citeseer, Pubmed, Actor, Squirrel, Chameleon, ogbn-proteins, ogbn-arxiv, ogbn-products and ogbn-papers100M. For each dataset, we use the same train/validation/test splits and evaluation metrics as [82]. Arxiv-year is a citation network among all computer science arxiv papers, as described by [55]. We use the public splits shared by [55], with a train/validation/test split ratio of 50%/25%/25%.
Hardware Specification Yes Our implementation is based on Py G [25] and DGL [76]. The experiments are conducted on a single workstation with 4 RTX 3090 GPUs and a quad-core CPU.
Software Dependencies No The paper mentions software like "Py G [25] and DGL [76]" but does not specify their version numbers, which is necessary for a reproducible description of ancillary software.
Experiment Setup Yes For fair comparisons, we use the same hyperparameters (including the number of layers, hidden dimension etc.), PE and readout as the baseline transformers. For GOAT + HDSE, Tables 9, 10 and 11 showcase the corresponding hyperparameters and coarsening algorithms. We fix the maximum distance length L = 30 and vary the maximum hierarchy level K in {0, 1, 2} in all experiments.