Efficient Graph Continual Learning via Lightweight Graph Neural Tangent Kernels-based Dataset Distillation
Authors: Rihong Qiu, Xinke Jiang, Yuchen Fang, Hongbin Lai, Hao Miao, Xu Chu, Junfeng Zhao, Yasha Wang
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments on several datasets show that LIGHTGNTK achieves state-of-the-art performance in GCL scenarios, promoting the development of adaptive and scalable LGMs. |
| Researcher Affiliation | Academia | 1School of Computer Science, Peking University, Beijing, China 2Key Laboratory of High Confidence Software Technologies, Ministry of Education, Beijing, China 3University of Electronic Science and Technology of China 4Department of Computer Science, Aalborg University 5Center on Frontiers of Computing Studies, Peking University, Beijing, China 6Peking University Information Technology Institute, Tianjin Binhai, China 7Big Data Technology Research Center, Nanhu Laboratory, Jiaxing, China 8National Engineering Research Center For Software Engineering, Peking University, Beijing, China. Correspondence to: Junfeng Zhao, Yasha Wang <Rihong EMAIL>. |
| Pseudocode | No | The paper describes methods in text and mathematical formulations but does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | The code and datasets are available at https://github.com/Artessay/Light GNTK. |
| Open Datasets | Yes | We conduct a comprehensive evaluation using a total of 13 datasets across three types of graph learning tasks. For graph classification, we utilize 7 datasets: NCI1, NCI109, PROTEINS, and DD from TUDataset (Morris et al., 2020), as well as ogbg-molhiv, ogbg-molbbbp, and ogbg-molbace from Open Graph Benchmarks (OGB) (Hu et al., 2020a). For node classification, we employ 4 citation network datasets: Cora, Cite Seer, and Pub Med from the Planetoid Dataset (Yang et al., 2016), along with ogbn-arxiv from OGB. For link prediction, we include 2 datasets: ogbl-collab and ogbl-ddi, both from OGB. Further details for all datasets are provided in Table D.1 of Appendix D.1. The code and datasets are available at https://github.com/Artessay/Light GNTK. |
| Dataset Splits | Yes | For the graph classification tasks in TUDataset, we split each dataset into 80%, 10%, and 10% for training, validation, and test sets, respectively, and use accuracy (ACC) as the evaluation metric. For the Planetoid and OGB datasets, we follow the official splits and corresponding evaluation metrics. |
| Hardware Specification | Yes | All implementations are carried out using Py Torch 2.5.1 with Python 3.12 on NVIDIA Ge Force RTX 3090 GPUs. |
| Software Dependencies | Yes | All implementations are carried out using Py Torch 2.5.1 with Python 3.12 on NVIDIA Ge Force RTX 3090 GPUs. |
| Experiment Setup | Yes | To ensure reproducibility, we optimize the parameters of baseline models using the Adam Optimizer with L2 regularization. During the pretrain process of the GNN backbone, a binary classification task is employed to distinguish between positive and negative edge samples on graphs. For the downstream tasks, a classification head is added above GNN backbone to predict. Furthermore, An early-stopping strategy with a patience of 50 epochs is employed to mitigate overfitting. Throughout all experiments, we use Graph Isomorphism Network (GIN) (Xu et al., 2019) as the GNN backbone for our LGMs. Based on empirical observations, we set the sampling rate to 0.1 in experiments, which offers a favorable trade-off between performance and computational efficiency. |