Use this colab notebook to reproduce our results

If you use this code or part of it, please cite the original reference:

Minici, M., Cinus, F., Bonchi, F., & Manco, G. (2024, October). Link Polarity Prediction from Sparse and Noisy Labels. In Proceedings of the 33rd ACM International Conference on Information & Knowledge Management. doi: https://doi.org/10.1145/3511808.3557253

---

Data Preprocessing

• Run python pre-processing-unsupervised-experiment.py for each dataset (bitcoin_alpha, bitcoin_otc, wiki, slashdot) and noise percentage (0.0, 0.1, 0.2). Alternatively, you can use the preprocessed files we will soon update on Zenodo.

---

Running scripts

Each dataset has a different set of hyperparameters, change their values accordingly in the run_SDGNN_lrw_MicroMesoSB.py.

• bitcoin_alpha:
  • unlabeled_perc = [None, ]
  • init_eps_one = True
• bitcoin_otc:
  • unlabeled_perc = [0.8, ]
  • init_eps_one = True
• wiki:
  • unlabeled_perc = [None, ]
  • init_eps_one = True
• slashdot:
  • unlabeled_perc = [0.5, ]
  • init_eps_one = False

We ensure other researchers can reproduce our results using this ready-to-use colab notebook.

For the sake of experimentation velocity, we will soon update our preprocessed files on a Zenodo node. However, you can preprocess the files by yourself using the pre-processing-unsupervised-experiment.py script (present in the Colab environment).