Pairwise Alignment Improves Graph Domain Adaptation
Authors: Shikun Liu, Deyu Zou, Han Zhao, Pan Li
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method demonstrates superior performance in real-world applications, including node classification with region shift in social networks, and the pileup mitigation task in particle colliding experiments. For the first application, we also curate the largest dataset by far for GDA studies. Our method shows strong performance in synthetic and other existing benchmark datasets. |
| Researcher Affiliation | Academia | 1Department of Electrical and Computer Engineering, Georgia Institute of Technology, Georgia, USA 2School of Data Science, University of Science and Technology of China, Hefei, China 3Department of Computer Science, University of Illinois Urbana Champaign, Champaign, USA. |
| Pseudocode | Yes | Algorithm 1 Pairwise Alignment |
| Open Source Code | Yes | Our code and data are available at: https://github. com/Graph-COM/Pair-Align |
| Open Datasets | Yes | To demonstrate the effectiveness of our pipeline, we curate the regional MAG data that partitions large citation networks according to the regions where papers got published (Hu et al., 2020; Wang et al., 2020)...Pileup Mitigation (Liu et al., 2023) is a dataset of a denoising task...Arxiv (Hu et et al., 2020) is a citation network...DBLP and ACM (Tang et al., 2008; Wu et al., 2020) are two paper citation networks... |
| Dataset Splits | Yes | The source graph is used for training, 20 percent of the node labels in the target graph are used for validation and the rest 80 percent are held out for testing. |
| Hardware Specification | Yes | All experiments are run on NVIDIA RTX A6000 with 48G memory and Quadro RTX 6000 with 24G memory. |
| Software Dependencies | No | The paper mentions software components and settings like 'Graph SAGE as backbones' and a 'learning rate of 0.003' but does not specify software versions (e.g., PyTorch 1.x, Python 3.x, CUDA x.x) for reproducibility. |
| Experiment Setup | Yes | The learning rate is 0.003 and the number of epochs is 400 for all experiments. The hidden dimension for GNN is 300 for Arxiv and MAG, 50 for Pileup, 128 for the DBLP/ACM dataset and 20 for synthetic datasets. The classifier dimension 300 for Arxiv and MAG, 50 for Pileup, 40 for DBLP/ACM dataset and 20 for synthetic datasets. The hyperparameters are tuned mainly for the robustness control, as the δ in regularizing edges and λ in L2 regularization for optimization of w and β. |