HyperIM

PyTorch implementation of Hyperbolic Interaction Model For Hierarchical Multi-Label Classification

Requirements

• torch>=1.0.0
• geoopt ($ pip install git+https://github.com/geoopt/geoopt.git)
• numpy
• scipy
• pandas
• tqdm

Instruction

Run HyperIM via

$ python HyperIM.py

or run EuclideanIM via

$ python EuclideanIM.py

Alternatively use the two Jupyter notebooks.

Data

X_train and X_test should be dense numpy array with shape (instance_num, word_num), y_train and y_test should be one-hot sparse scipy array with shape (instance_num, label_num). Sample data is provided in ./data/sample/.

The multi-label text classification datasets equipped with hierarchically structured labels (RCV1, Zhihu and WikiLSHTC) are publicly available.

Pre-trained embeddings

Hyperbolic word embeddings can be trained following Poincaré GloVe. Pre-trained word embeddings should have shape (vocab_size, embed_dim).

The label hierarchy can be embedded using the gensim implementation of the Poincaré embeddings, specified in Train and use Poincaré embeddings. Label embeddings should have shape (label_num, embed_dim). Note that the index of labels in the label embeddings should be consistent with y_train and y_test.

Use them accordingly in HyperIM.py and EuclideanIM.py.

Citation

If you find this code useful for your research, please cite the following paper in your publication:

@article{chen2019hyperbolic,
  title={Hyperbolic Interaction Model For Hierarchical Multi-Label Classification},
  author={Chen, Boli and Huang, Xin and Xiao, Lin and Cai, Zixin and Jing, Liping },
  journal={arXiv preprint arXiv:1905.10802},
  year={2019}
}