Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Rethinking Tokenized Graph Transformers for Node Classification
Authors: Jinsong Chen, Chenyang Li, Gaichao Li, John E. Hopcroft, Kun He
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive empirical results on various datasets showcase the superiority of Swap GT for node classification. 5 Experiments |
| Researcher Affiliation | Academia | 1School of Computer Science and Technology, Huazhong University of Science and Technology 2Faculty of Artificial Intelligence in Education, Central China Normal University 3Hopcroft Center on Computing Science, Huazhong University of Science and Technology 4Department of Computer Science, Cornell University EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 The Token Swapping Algorithm |
| Open Source Code | Yes | Code is available at https://github.com/JHL-HUST/Swap GT. |
| Open Datasets | Yes | We adopt eight widely used datasets, involving homophily and heterophily graphs: Photo [6], ACM [34], Computer [6], Blog Catalog [29], UAI2010 [35], Flickr [29] and Wiki-CS [31]. |
| Dataset Splits | Yes | In dense splitting, we randomly choose 50% of each label as the training set, 25% as the validation set, and the rest as the test set... While in sparse splitting [15], we adopt 2.5%/2.5%/95% splitting for training set, validation set and test set. |
| Hardware Specification | Yes | All experiments are implemented using Python 3.8, Py Torch 1.11, and CUDA 11.0 and executed on a Linux server with an Intel Xeon Silver 4210 processor, 256 GB of RAM, and a 2080TI GPU. |
| Software Dependencies | Yes | All experiments are implemented using Python 3.8, Py Torch 1.11, and CUDA 11.0 |
| Experiment Setup | Yes | For Swap GT, we employ a grid search strategy to identify the optimal parameter settings. Specifically, We try the learning rate in {0.001, 0.005, 0.01}, dropout in {0.3, 0.5, 0.7}, dimension of hidden representations in {256, 512}, k in {4, 6, 8}, α in {0.1, . . . , 0.9}. |