Does Invariant Graph Learning via Environment Augmentation Learn Invariance?
Authors: Yongqiang Chen, Yatao Bian, Kaiwen Zhou, Binghui Xie, Bo Han, James Cheng
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on 12 datasets including Drug OOD with various graph distribution shifts confirm the effectiveness of GALA. We conduct extensive experiments to validate the effectiveness of GALA using 12 datasets with various graph distribution shifts. |
| Researcher Affiliation | Collaboration | Yongqiang Chen1 , Yatao Bian2, Kaiwen Zhou1 1The Chinese University of Hong Kong 2Tencent AI Lab {yqchen,kwzhou}@cse.cuhk.edu.hk yatao.bian@gmail.com Binghui Xie1, Bo Han3, James Cheng1 3Hong Kong Baptist University bhanml@comp.hkbu.edu.hk {bhxie21,jcheng}@cse.cuhk.edu.hk |
| Pseudocode | Yes | Algorithm 1 GALA: Graph inv Ariant Learning Assistant and Algorithm 2 GALA: Clustering based Graph inv Ariant Learning Assistant |
| Open Source Code | Yes | 1Code is available at https://github.com/LFhase/GALA. |
| Open Datasets | Yes | We adopt BA-2motifs [50] to implement 4 variants of 3-class two-piece graph (Def. 3.1) datasets. We also adopt datasets containing various realistic graph distribution shifts to comprehensively evaluate the OOD performance of GALA. We adopt 6 datasets from Drug OOD benchmark [28], which focuses on the challenging real-world task of AI-aided drug affinity prediction. The Drug OOD datasets include splits using Assay, Scaffold, and Size from the EC50 category (denoted as EC50-*) and the Ki category (denoted as Ki-*). We also adopt graphs converted from the Colored MNIST dataset [3] using the algorithm from Knyazev et al. [36], which contains distribution shifts in node attributes (denoted as CMNIST-sp). In addition, we adopt Graph-SST2 [84] |
| Dataset Splits | Yes | Table 5: Information about the datasets used in experiments. ... # Training # Validation # Testing ... (e.g., Two-piece graphs {0.8, 0.6} 9, 000 3, 000 3, 000) |
| Hardware Specification | Yes | We ran our experiments on Linux Servers installed with V100 graphics cards and CUDA 10.2. |
| Software Dependencies | No | We implement our methods with Py Torch [59] and Py Torch Geometric [18]. We ran our experiments on Linux Servers installed with V100 graphics cards and CUDA 10.2. It mentions PyTorch and PyTorch Geometric but does not specify their version numbers, only CUDA 10.2 has a version. |
| Experiment Setup | Yes | By default, we use Adam optimizer [34] with a learning rate of 1e 3 and a batch size of 128 for all models at all datasets. Except for CMNIST-sp, we use a batch size of 256 to facilitate the evaluation following previous works [55]. We use 3-layer GIN [75] with Batch Normalization [27] between layers and JK residual connections at the last layer [74]. The hidden dimension is set to 32 for Two-piece graphs, CMNIST-sp, and 128 for SST2, and Drug OOD datasets. The pooling is by default a mean function over all nodes. The only exception is Drug OOD datasets, where we follow the backbone used in the paper [28], i.e., 4-layer GIN with sum readout. |