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..
GraphKeeper: Graph Domain-Incremental Learning via Knowledge Disentanglement and Preservation
Authors: Zihao Guo, Qingyun Sun, Ziwei Zhang, Haonan Yuan, HUIPING ZHUANG, Xingcheng Fu, Jianxin Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the proposed Graph Keeper achieves state-ofthe-art results with 6.5% 16.6% improvement over the runner-up with negligible forgetting. Moreover, we show Graph Keeper can be seamlessly integrated with various representative GFMs, highlighting its broad applicative potential. |
| Researcher Affiliation | Academia | Zihao Guo1, Qingyun Sun1 , Ziwei Zhang1, Haonan Yuan1, Huiping Zhuang2, Xingcheng Fu3, Jianxin Li1 1SKLCCSE, School of Computer Science and Engineering, Beihang University 2Shien-Ming Wu School of Intelligent Engineering, South China University of Technology 3Key Lab of Education Blockchain and Intelligent Technology, Guangxi Normal University EMAIL EMAIL, EMAIL |
| Pseudocode | Yes | The overall training and inference process of Graph Keeper are given in Algorithm 1 and Algorithm 2. Algorithm 1: The overall training process of Graph Keeper Algorithm 2: The inference process of Graph Keeper |
| Open Source Code | Yes | 2https://github.com/Ring BDStack/Graph Keeper The code and datasets is available at: https://anonymous.4open. science/r/Graph Keeper. |
| Open Datasets | Yes | To evaluate Graph Keeper2, we conduct comprehensive experiments on 15 real-world datasets, where detailed descriptions are in Appendix C.1. ... We conduct experiments on 15 real-world datasets, including academic networks (Cora [26], Citeseer [12], Pub Med [21], Coauthor CS [27], and DBLP [7]), co-purchase networks (Photo [27] and Computer [27]), web networks (Wiki CS [20], Facebook [24], Chameleon [24], and Squirrel [24]), social networks (Git Hub [24], Last FMAsia [25], and Deezer Europe [25]), and airline networks (Airport [3]). |
| Dataset Splits | Yes | For each graph domain, we set the unified dimension of the features to 512 and split the training set, validation set, and test set in proportions of 60%, 20%, and 20%. |
| Hardware Specification | Yes | Operating System: Ubuntu 20.04 LTS. CPU: Intel(R) Xeon(R) Platinum 8358 CPU@2.60GHz with 1TB DDR4 of Memory. GPU: NVIDIA Tesla V100 with 32GB of Memory. |
| Software Dependencies | Yes | Software: CUDA 11.7, Python 3.8.0, Pytorch 1.7.1, DGL 0.6.1. |
| Experiment Setup | Yes | We set the number of model layers to 2 for all methods, with the learning rate set to 5e-2, the weight decay coefficient set to 5e-4, and 200 training epochs per incremental graph domain. The parameters of models are optimized by Adam [11]. For Graph Keeper, the pre-trained GNN model is trained through link prediction on the first incremental domain (except for Sec. 5.3), the feature and structure augmentation are achieved by randomly masking a few features and dropping a few edges, and the embedding prototype sampling is implemented through DBSCAN [5]. The hyperparameter r is chosen from {4, 8, 16, 32, 64, 128}, γ1 is chosen from {0.01, 0.1, 1.0, 5.0, 10.0}, γ2 is chosen from {0.001, 0.01, 0.10, 0.50, 1.0}. |