GFT: Graph Foundation Model with Transferable Tree Vocabulary
Authors: Zehong Wang, Zheyuan Zhang, Nitesh Chawla, Chuxu Zhang, Yanfang Ye
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results demonstrate the effectiveness of GFT in graph learning on cross-task and cross-domain datasets. and 4 Experiments |
| Researcher Affiliation | Academia | Zehong Wang University of Notre Dame Indiana, USA zwang43@nd.edu Zheyuan Zhang University of Notre Dame Indiana, USA zzhang42@nd.edu Nitesh V Chawla University of Notre Dame Indiana, USA nchawla@nd.edu Chuxu Zhang University of Connecticut Connecticut, USA chuxu.zhang@uconn.edu Yanfang Ye University of Notre Dame Indiana, USA yye7@nd.edu |
| Pseudocode | No | The paper describes its methods in prose and equations (e.g., Section 3, Section 3.1), but it does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The open source code and data are available at https://github.com/Zehong-Wang/GFT. |
| Open Datasets | Yes | For node-level tasks, we utilize citation networks such as Cora, Pub Med, Arxiv, and the web link network Wiki CS. For edge-level tasks, we include two Knowledge Graphs (KGs), WN18RR and FB15K237. For graph-level tasks, we use molecule networks, including HIV, PCBA, and Ch EMBL. All preprocessing steps follow [45]. |
| Dataset Splits | Yes | For Cora and Pub Med, we follow the common split setting with 20 labeled nodes per class for training, utilizing a predefined 10 splits with different seeds to report average performance. For Wiki CS, we also employ the standard split, reporting average accuracy across 20 different training splits, each with 20 random seeds, and using 5% of nodes in each class for training. For Arxiv, HIV, and PCBA, we follow the official splits, conducting experiments 10 times with random seeds to determine average accuracy. For WN18RR and FB15K237, we follow the splits outlined in Liu et al. [45]. Specifically, for FB15K237, the training set comprises 272,115 edges, the validation set 17,535 edges, and the test set 20,466 edges; for WN18RR, the numbers are 86,835, 3,034, and 3,134, respectively. |
| Hardware Specification | Yes | We utilize an NVIDIA A40 with 48GB GPU memory for all experiments. Both the pre-training and fine-tuning phases can be conducted on a single Nvidia Ge Force RTX 3090 with 24GB memory. |
| Software Dependencies | No | The paper mentions software components like Graph SAGE-like architecture, Sentence Transformer [60], and Adam W optimizer, but it does not specify their version numbers (e.g., 'using the Sentence Transformer [60]' without a version for the library). |
| Experiment Setup | Yes | Appendix F.4 'Hyper-parameter Setting' and Table 11 provide extensive details on hyperparameters such as learning rates, number of epochs, batch sizes, weight decays, hidden dimensions, activation functions, and specific loss weights (e.g., 'The pre-training phase lasts for 25 epochs with a batch size of 1024.'). |