This repo contains official implementation for HyLa (HYperbolic LAplacian features), described in the paper Laplacian Features for Learning with Hyperbolic Space.
Our HyLa feature, when used together with a simple linear graph model, i.e., SGC, can already achieves impressive performance. For example, on the node classification tasks over standard benchmarks with 16 dimensional hypebolic embeddings (baseline results taken from corresponding papers):
| Dataset (Hyperbolicity) | HyLa-SGC | HGCN | SGC |
|---|---|---|---|
| Disease (0.0) | 91.3 % | 74.5 % | 69.5% |
| Airport (1.0) | 95.8% | 90.6 % | 80.6% |
| Pubmed (3.5) | 81.1 % | 80.3 % | 78.9% |
| Citeseer (5.0) | 72 % | 64.0 % | 71.9% |
| Cora (11.0) | 82.6 % | 79.9 % | 81.0% |
PyTorch>=1.0.0 is required to run the implementation, other dependences are provided in requirements.txt, before training, set up environments by simply running
$ conda env create -f environment.yml
$ conda activate hyla
This repo contains implementation for both node classification (./nc) and text classification (./text) task. For the node classification task, The ./nc/datasets/ folder includes both standard benchmarks (Cora, Citeseer, and Pubmed), social network (Reddit) and hyperbolid datasets (Airport, and Disease). Due to space limit, please download reddit dataset from FastGCN and put reddit_adj.npz, reddit.npz under ./nc/datasets/reddit/.
Implementation for text classification following (Graph Convolutional Networks for Text Classification and Simplifying Graph Convolutional Networks).
Code modified from repo Text-GCN(https://github.com/yao8839836/text_gcn) and SGC(https://github.com/Tiiiger/SGC). Both transductive and inductive settings are considered, where HyLa-SGC outperforms Text-GCN and Text-SGC especially in the inductive setting. We get the text data (R8, R52, Ohsumed, and MR) from Text-GCN repo and process with remove_words.py, build_graph.py following SGC repo with some slight changes to the code. If you want to redo the processing, see options with python build_graph.py --help and python remove_words.py --help. Please put processed data into ./text/datasets/processed_data/.
We tune the hyperparameters with grid search for all datasets at both 16 (./nc/hyper_parameters_16d.json) and 50 dimension (./nc/hyper_parameters_50d.json). Simply specify the dataset and dimension in train-nc.sh and run
$ bash train-nc.sh
to get results in the paper. particularly, please remove -use_feats option for airport dataset and include -inductive option for inductive training on reddit. For a detailed look of the code, see more options with python ./nc/HyLa.py --help and add hyperparameters you want to change in train-nc.sh.
We tune the hyperparameters with grid search for all datasets at both transductive (./text/hyper_parameters_transductive.json) and inductive (./text/hyper_parameters_inductive.json) setting. The hyperbolic embedding dimension is fixed at 50 for consistency. Please specify training options (e.g. datasets) in train-text.sh and run
$ bash train-text.sh
to get results in the paper. See more options with python ./text/TextHyLa.py --help and add extra hyperparameters in train-text.sh to play with it.
Early stopping is provided in the code, i.e., resume back to the model with best val accuracy, though it's not used when reporting results. Some of the code was also adapted from poincare-embeddings and hgcn to set up e.g., baselines, hyperbolic models and optimizers.
If you find this repo useful, please cite:
@article{yu2022hyla,
title={Laplacian Features for Learning with Hyperbolic Space},
author={Yu, Tao and De Sa, Christopher},
journal={arXiv preprint arXiv:2202.06854},
year={2022}
}