Unified Graph Neural Networks Pre-training for Multi-domain Graphs

Authors: Mingkai Lin, Xiaobin Hong, Wenzhong Li, Sanglu Lu

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the effectiveness of MDPGNN through theoretical analysis and extensive experiments on four real-world graph datasets, showing its superiority in enhancing GNN performance across diverse domains.
Researcher Affiliation Academia State Key Laboratory for Novel Software Technology, Nanjing University Nanjing, China EMAIL, EMAIL
Pseudocode Yes Algorithm 1: Pre-training Process for MDP-GNN
Open Source Code No The paper does not explicitly state that source code for the described methodology is publicly available or provide a link to a repository.
Open Datasets Yes We evaluate MDP-GNN using four large-scale text-free graphs from distinct domains: (1) Academic (Hu et al. 2020a), a citation network with papers indexed by MAG (Wang et al. 2020); (2) Product (Hu et al. 2020a), an Amazon product co-purchasing network; (3) Reddit (Hamilton, Ying, and Leskovec 2017), a comment graph derived from Reddit; and (4) Yelp (Zeng et al. 2019), a social network formed from the Yelp platform.
Dataset Splits Yes For the testing graphs, we allocate up to 10% of the known labels for training and another 10% for validation.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments.
Software Dependencies No The paper mentions a 'two-layer GT' as the GNN backbone and the 'Adam optimizer', but it does not specify version numbers for any software libraries or frameworks used.
Experiment Setup Yes We tune all the models for 1000 epochs using the Adam optimizer, with a learning rate of 0.005. Each experiment is conducted five times, and the mean results are reported. The conversion functions {ϕk} for feature integration are implemented with MLPs. For the bi-level connection learning of Eq. 4, the inner and outer loops are both set to 10 with a learning rate of 0.001, as well as for the inner step of in Eq. 6. The rest hyper-parameters are set as τ = 0.8, λ1 = 1, λ2 = 0.01, Np = 256.