RSDNE (python light version)

RSDNE: Exploring Relaxed Similarity and Dissimilarity from Completely-imbalanced Labels for Network Embedding. AAAI18. This is a shallow method for the problem of Zero-shot Graph Embedding (ZGE), i.e., graph embeddings when labeled data cannot cover all classes.

• This is the python light version of RSNDE.
• The original code is written in matlab (code).
• More datasets can be found in (code).

Breifly explain:

• RSDNE loss: min_{U,V} J = |G-UV|^2 + alpha*( tr(U'LsU) + tr(U'LwU) ) + lambda(|U|^2 + |V|^2)

• By setting alpha=0 (i.e., removing the relax part), our method will reduce to the common matrix decomposition method, like MFDW (IJCAI15).

Usage (abstract):

• set the dataset
• python main_RSDNE.py

Experiment results:

• label rate 30%:

  • MFDW: (0.5724018973695558, 0.5283697422985723) # set alpha=0
  • RSDNE: (0.6018602846054334, 0.5766616831506641)

• label rate 50%:

  • MFDW: (0.6144927536231883, 0.5603593000330456) # set alpha=0
  • RSDNE: (0.6661835748792271, 0.6218905028673257)

Citing

If you find RSDNE useful in your research, please cit our paper, thx:

@InProceedings{wang2018rsdne,
  title={{RSDNE}: Exploring Relaxed Similarity and Dissimilarity from Completely-imbalanced Labels for Network Embedding},
  author={Wang, Zheng and Ye, Xiaojun and Wang, Chaokun and Wu, YueXin and Wang, Changping and Liang, Kaiwen},
  booktitle={AAAI},
  pages={475--482},
  year={2018}
}