Contextual Structure Knowledge Transfer for Graph Neural Networks
Authors: Zhiyuan Yu, Wenzhong Li, Zhangyue Yin, Xiaobin Hong, Shijian Xiao, Sanglu Lu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments conducted on six benchmark datasets substantiate the superiority of CSGNN over the state-of-the-art methods, demonstrating significant improvements in accuracy and robustness against homophily shifts. |
| Researcher Affiliation | Academia | 1State Key Laboratory for Novel Software Technology, Nanjing University 2School of Computer Science, Fudan University EMAIL, {yinzy21}@m.fudan.edu.cn, EMAIL |
| Pseudocode | No | The paper describes the methodology using textual explanations and mathematical formulations, but it does not include any clearly labeled pseudocode blocks or algorithms. |
| Open Source Code | Yes | The source code for CS-GNN is publicly available at https://github.com/yyy0959/CS-GNN. |
| Open Datasets | Yes | We conduct experiments on six real-world datasets: Citation (Tang et al. 2008), Social (Li et al. 2015), Web KB (Pei et al. 2020), Airport (Ribeiro, Saverese, and Figueiredo 2017), ARXIV (Hu et al. 2020), and CORA (Bojchevski and G unnemann 2017). |
| Dataset Splits | Yes | The results show that when the number of labeled nodes in the target domain is sparse (e.g., p = 0.1%), most existing methods struggle to perform well, resulting in lower accuracy. However, the proposed method, CS-GNN, demonstrates a significant improvement, achieving approximately 10% accuracy gain compared to the baseline models. As the percentage of labeled nodes increases, the performance of all methods increases accordingly. When the target domain has a sufficient number of labeled nodes (e.g., p = 10%), the knowledge from the source domain is overshadowed by fine-tuning on the target domain, leading to high classification performance across all methods. |
| Hardware Specification | No | The paper describes the experimental setup and results but does not provide specific details about the hardware used to run the experiments, such as CPU or GPU models. |
| Software Dependencies | No | For all compared methods, we set their depth to 2 layers and use the implementations from the Py Torch Geometric Library in all experiments. [...] Therefore we utilize the scalable graph partition module METIS (Karypis and Kumar 1998) to split nodes into multiple non-overlapping subgraphs. |
| Experiment Setup | Yes | For all compared methods, we set their depth to 2 layers and use the implementations from the Py Torch Geometric Library in all experiments. The representation dimension is set to 64. Other hyperparameters of baseline methods are set to the suggested values in their respective papers or are carefully tuned for fairness. ... We conducted a comprehensive hyperparameter analysis to understand the sensitivity and impact of key parameters in CS-GNN. Specifically, we examined the effects of four hyperparameters: the number of moments of feature smoothness considered (denoted by n) and the hop of the ego-networks (denoted by k), the number of network partitions (denoted by m) and temperature τ, as shown in Figure 4. ... A temperature of 10 provided the highest accuracy, with lower and higher values resulting in a performance drop. |