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 [1].

Towards Understanding Parametric Generalized Category Discovery on Graphs

Authors: Bowen Deng, Lele Fu, Jialong Chen, Sheng Huang, Tianchi Liao, Zhang Tao, Chuan Chen

ICML 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results validate our (theoretical) findings and demonstrate SWIRL s effectiveness. ... We validate all (theoretical) analyses on synthetic datasets and demonstrate SWIRL s effectiveness on real-world graphs.
Researcher Affiliation Academia 1School of Computer Science and Engineering, Sun Yat-Sen University, Guangzhou, China 2School of Systems Science and Engineering, Sun Yat-Sen University, Guangzhou, China 3School of Software Engineering, Sun Yat Sen University, Zhuhai, China. Correspondence to: Chuan Chen <EMAIL>.
Pseudocode Yes Algorithm 1 The full procedure of SWIRL
Open Source Code No The paper does not explicitly state that the source code for the methodology described is open-source or provide a link to a repository. It discusses third-party tools and benchmarks but not its own implementation code.
Open Datasets Yes We created node-level GGCD datasets based on five existing datasets: Cora, Citeseer, Wiki, A-Computers, and A-Photo. For Cora and Citeseer, we used the public splits, while for the other three datasets, The entire node set is stratified into train, validation, and test subsets in a 2:2:6 ratio.
Dataset Splits Yes The entire node set is stratified into train, validation, and test subsets in a 2:2:6 ratio.
Hardware Specification Yes The first system runs Ubuntu 22.04 and is equipped with an RTX 4090 GPU (24GB), an Intel i7-12700 CPU, and 64GB of RAM. The second system, which uses Ubuntu 20.04, features an RTX 4090 GPU (24GB), dual Intel Xeon Gold 6240C processors, and 126GB of RAM.
Software Dependencies Yes Both systems have the same Conda environment, which includes Py Torch 2.5 (Paszke et al., 2017) and Py G 2.5 (Fey & Lenssen, 2019), all built on CUDA 12.1.
Experiment Setup Yes We train the model for 1000 epochs using the Adam optimizer (Kingma & Ba, 2017) with a learning rate of 0.01. ... The training loss is LSW = (1 α2)(LNCE + β1LSW + α1LER) + α2LCE. We set α2 = 0.35 and α1 = 2, which are commonly used values in many baseline models (Vaze et al., 2022; Wen et al., 2023), and choose β1 = 20. ... Table 5: Hyperparameters of GGCD methods and the corresponding values or search spaces