DINE: Dimensional Interpretability of Node Embeddings

In this repository you will find code to run experiments of the paper DINE: Dimensional Interpretability of Node Embeddings. The code has been tested with Python 3.6.13.

If you use the code in this repository, please consider citing us:

@article{piaggesi2023dine,
  title={DINE: Dimensional Interpretability of Node Embeddings},
  author={Piaggesi, Simone and Khosla, Megha and Panisson, Andr{\'e} and Anand, Avishek},
  journal={arXiv preprint arXiv:2310.01162},
  year={2023}
}

Repository organization

Python scripts

To generate DINE embedding, run the following command:

python3 dine.py --input input_embedding \
                --output output_embedding \
                --emb-dim size \
                --noise-level 0.2 \
                --num-epochs 2000 \
                --learning-rate 0.1 \
                --lambda-size l_size \
                --lambda-orth l_orth \

The input and output embedding are compressed text files where each line contains the node index and embedding entries in a space-separated format:

node0 0.21 0.14 0.48 ...
node1 0.35 0.56 0.12 ...
node2 ...

Input embeddings can be generated by any user-defined method. Please refer to deepwalk.py to write your own embedding script.

Jupyter notebooks

• DeepWalk-Subgraphs-KarateClub.ipynb It shows how to draw pictures of utility-induced subgraphs on the Karate Club graph.

• DeepWalk+DINE-Interpretability-Scores-SBM.ipynb It shows how to compare interpretability metrics on DeepWalk embeddings before and after the application of DINE, on synthetic graph generated with the Stochastic Block Model.

• DeepWalk+DINE-Link-Prediction-Cora.ipynb It shows the link prediction performance of DeepWalk embeddings before and after the application of DINE on the Cora graph.