Environment-Aware Dynamic Graph Learning for Out-of-Distribution Generalization
Authors: Haonan Yuan, Qingyun Sun, Xingcheng Fu, Ziwei Zhang, Cheng Ji, Hao Peng, Jianxin Li
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on real-world and synthetic dynamic graph datasets demonstrate the superiority of our method against state-of-the-art baselines under distribution shifts. |
| Researcher Affiliation | Academia | 1Beijing Advanced Innovation Center for Big Data and Brain Computing, Beihang University 2School of Computer Science and Engineering, Beihang University 3Key Lab of Education Blockchain and Intelligent Technology, Guangxi Normal University 4Department of Computer Science and Technology, Tsinghua University |
| Pseudocode | Yes | Algorithm 1: Overall training process of EAGLE. Input: Dynamic graph DG = ({G}T t=1) with labels Y1:T of link occurrence; Number of training epochs E; Number of intervention times S; Hyperparameters α and β. Output: Optimized model f θ; Predicted label Y T of link occurrence at time T + 1. |
| Open Source Code | Yes | Code is available at https://github.com/Ring BDStack/EAGLE on Py Torch [63] and Mind Spore [33]. |
| Open Datasets | Yes | We use three real-world datasets to evaluate EAGLE1 on the challenging link prediction task. COLLAB [81] is an academic collaboration dataset... Yelp [75] contains customer reviews on business... ACT [45] describes students actions on a MOOC platform... |
| Dataset Splits | Yes | We use 10/1/5 chronological graph snapshots for training, validation, and testing, respectively. |
| Hardware Specification | Yes | CPU: Intel(R) Xeon(R) Platinum 8358 CPU@2.60GHz with 1TB DDR4 of Memory. GPU: NVIDIA Tesla A100 SMX4 with 40GB of Memory. |
| Software Dependencies | Yes | Software: CUDA 10.1, Python 3.8.12, Py Torch [63] 1.9.1, Py Torch Geometric [20] 2.0.1. |
| Experiment Setup | Yes | For our EAGLE, the hyperparameter α is chosen from {10 3, 10 2, 10 1, 100, 101}, and β is chosen from {10 6, 10 5, 10 4, 10 3, 10 2}. The intervention ratio and the mixing ratio are carefully tuned for each dataset. For other parameters, we adopt the Adam optimizer [40] with an appropriate learning rate and weight decay for each dataset and adopt the grid search for the best performance using the validation split. |