Learning Invariant Representations of Graph Neural Networks via Cluster Generalization
Authors: Donglin Xia, Xiao Wang, Nian Liu, Chuan Shi
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present experiments to explore this question. We generate graph structure and node features. Subsequently we train GNNs on the initially generated graph structure and gradually change the structures to test the generalization of GNNs (more details are in Section 2).Our proposed mechanism is a friendly plug-in, which can be easily used to improve most of the current GNNs. We conduct extensive experiments on three typical structure shift tasks. The results well demonstrate that our proposed model consistently improves generalization ability of GNNs on structure shift. |
| Researcher Affiliation | Academia | Donglin Xia1, Xiao Wang2 , Nian Liu1, Chuan Shi1 1Beijing University of Posts and Telecommunications 2Beihang University {donglin.xia, nianliu, shichuan}@bupt.edu.cn, xiao_wang@buaa.edu.cn |
| Pseudocode | Yes | the whole algorithm is shown in A.1 |
| Open Source Code | Yes | Code available at https://github.com/BUPT-GAMMA/CITGNN |
| Open Datasets | Yes | To comprehensively evaluate the proposed CIT mechanism, we use six diverse graph datasets. Cora, Citeseer, Pubmed [20], ACM, IMDB [27] and Twitch-Explicit [17]. |
| Dataset Splits | Yes | For Cora, Citeseer and Pubmed, we follow the original node-classification settings [13]... For Twitch-Explicit, we use six different networks. We take one network to train, one network to validate and the rest of networks to test. |
| Hardware Specification | No | The paper does not specify any hardware details (e.g., specific GPU or CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | We randomly sample 20 nodes per class for training... The probability of transfer p determines how many nodes to transfer every time. We vary its value... We make transfer process every k epochs. We vary k from 1 to 50... The number of clusters is the most important parameter in the clustering process. We vary m from 20 to 100 unevenly. |