GCAL: Adapting Graph Models to Evolving Domain Shifts
Authors: Ziyue Qiao, Qianyi Cai, Hao Dong, Jiawei Gu, Pengyang Wang, Meng Xiao, Xiao Luo, Hui Xiong
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental evaluations demonstrate that GCAL substantially outperforms existing methods in terms of adaptability and knowledge retention. We conduct extensive experiments on various graph datasets, demonstrating that GCAL significantly outperforms state-of-the-art across domain shifts. Section 4, titled 'Experiments', includes subsections such as 'Experimental Setup', 'Overall Performance Comparison', 'In-depth Analysis of Continuous Performance', 'Ablation Studies', and 'Hyperparameter Experiments'. |
| Researcher Affiliation | Academia | 1School of Computing and Information Technology, Great Bay University 2Thrust of Artificial Intelligence, The Hong Kong University of Science and Technology (Guangzhou) 3Computer Network Information Center, University of the Chinese Academy of Sciences 4University of Macau 5Department of Computer Science, University of California, Los Angeles 6Department of Computer Science and Engineering, The Hong Kong University of Science and Technology Hong Kong SAR, China. All listed affiliations are academic institutions. |
| Pseudocode | No | The paper describes the methodology in prose, detailing steps and equations within the text, but does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | Yes | The code of GCAL is available at https://github.com/joe817/GCAL. |
| Open Datasets | Yes | Our paper involves two primary categories of graph datasets, differentiated by regional and temporal shifts. For regional shifts, Facebook-100(Traud et al., 2012) and Twitch-Explicit(Rozemberczki et al., 2021) datasets consist of multiple social networks from different regions. For temporal shifts, OGB-Arxiv(Hu et al., 2020) is a paper citation network dataset, and Elliptic(Pareja et al., 2020) is a Bitcoin transactions network dataset, both of which include graphs from different time steps. |
| Dataset Splits | Yes | The training, validation, and test rates for pretrain datasets are set at 60%, 20%, and 20% respectively. For academic and social networks, initial training is conducted on three university graphs from the Facebook-100 dataset... For Twitch-explicit, the model is initially trained on the DE network and later tested for adaptability across regional networks... In addressing temporal shifts, the OGB-Arxiv dataset is employed, using data prior to 2011 to pretrain the model, while data from 2011 and later is used for continual adaptation... For the Elliptic dataset, we utilize snapshots 7 through 9 for initial training... with the remaining data employed for continual adaptation. |
| Hardware Specification | No | The computational resources are supported by Song Shan Lake HPC Center (SSL-HPC) in Great Bay University. This mentions a High-Performance Computing center but does not provide specific details on GPU models, CPU types, or memory used. |
| Software Dependencies | No | We use a 2-layer GCN as the backbone for three datasets, except for OGB-Arxiv, where we use Graph SAGE (Hamilton et al., 2017). The paper mentions specific graph neural network models used as backbones but does not provide version numbers for any specific software libraries (e.g., PyTorch, TensorFlow, CUDA) or other dependencies. |
| Experiment Setup | Yes | We use a 2-layer GCN as the backbone for three datasets, except for OGB-Arxiv, where we use Graph SAGE (Hamilton et al., 2017). During the train and adapt periods, the learning rates and weight decays are set as follows: lr = 0.0001 and wd = 5 10 4 for training, and lr = 0.001 and wd = 5 10 4 for adaptation. The number of epochs for pre-training is set between 100 and 200, while the adaptation phase involves a relatively smaller number of epochs, ranging from 1 to 10, for these four datasets. |