OpenIMR

The PyTorch implementation of OpenIMR.

OpenIMR is proposed for open-world semi-supervised learning for node classification (ICDE 2024).

Overview

You can enter the folder run/ and then run run_ours.sh (run_ours_large.sh) for open-world semi-supervised learning on the Coauthor CS dataset (ogbn-arxiv dataset).

Specifically, the repository is organized as follows:

• losses/ contains the implementation of supervised contrastive loss, which can be used for implementing the proposed PLCL loss.

• networks/ contains the implementation of a GAT backbone.

• networks_large/ contains the implementation of a GAT backbone for large graph datasets.

• util.py is used for loading and pre-processing the dataset, and also includes the functions for computing metrics.

• train_ours.py is used for implementing the pipeline of OpenIMR.

• train_ours_large.py is used for implementing the pipeline of OpenIMR for large graph datasets.

Requirements

Before running the code, you should install PyTorch, dgl, scipy, sklearn, numpy, and ogb.

Running the code

To run OpenIMA on the example dataset Coauthor CS

$ cd run/
$ sh run_ours.sh

To run OpenIMA on the larger example dataset ogbn-arxiv

$ cd run/
$ sh run_ours_large.sh

All the experiments are repeated ten times under ten different data splits, and the reported accuracy is averaged over the ten runs.

Reference

If you follow our idea in your work, please cite the following paper:

 @inproceedings{Wang2024OpenIMR,
     author = {Yanling Wang and Jing Zhang and Lingxi Zhang and Lixin Liu and Yuxiao Dong and Cuiping Li and Hong Chen and Hongzhi Yin},
     title = {Open-World Semi-Supervised Learning for Node Classification},
     booktitle = {ICDE},
     year = {2024}
   }