Handling Distribution Shifts on Graphs: An Invariance Perspective

Authors: Qitian Wu, Hengrui Zhang, Junchi Yan, David Wipf

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We prove the validity of our method by theoretically showing its guarantee of a valid OOD solution and further demonstrate its power on various real-world datasets for handling distribution shifts from artificial spurious features, cross-domain transfers and dynamic graph evolution.
Researcher Affiliation Collaboration Qitian Wu Shanghai Jiao Tong University echo740@sjtu.edu.cn Hengrui Zhang University of Illinois at Chicago hzhan55@uic.edu Junchi Yan Shanghai Jiao Tong University yanjunchi@sjtu.edu.cn David Wipf Amazon daviwipf@amazon.com
Pseudocode Yes Algorithm 1: Stable Learning for OOD Generalization in Node-Level Prediction on Graphs.
Open Source Code Yes 1The implementation is public available at https://github.com/qitianwu/Graph OOD-EERM.
Open Datasets Yes We adopt two public social network datasets Twitch-Explicit and Facebook-100 collected by Lim et al. (2021).
Dataset Splits Yes We generate 10-fold graph data with distinct environment id s and use 1/1/8 of them for training/validation/testing.
Hardware Specification Yes Most of our experiments are run on Ge Force RTX 2080Ti with 11GB except some experiments requiring large GPU memory for which we adopt RTX 8000 with 48GB.
Software Dependencies Yes The configurations of our environments and packages are listed below: Ubuntu 16.04, Numpy 1.20.3, Py Torch 1.9.0, Py Torch Geometric 1.7.2
Experiment Setup Yes Other hyper-parameters are searched with grid search on validation dataset. The searching space are as follows: learning rate for GNN backbone αf {0.0001, 0.0002, 0.001, 0.005, 0.01}, learning rate for graph editers αg {0.0001, 0.001, 0.005, 0.01}, weight for combination β {0.2, 0.5, 1.0, 2.0, 3.0}, number of edge editing for each node s {1, 5, 10}, number of iterations for inner update before one-step outer update T {1, 5}.