Unsupervised Multiple Kernel Learning for Graphs via Ordinality Preservation

Published as a conference paper at ICLR 2025.

Dependencies

• Python 3.8+
• PyTorch 2.2+
• Numpy 1.24+

The folder structure is:

|-- .Ensemble_Kernel
    |-- README.md
    |-- cache
    |-- experiment
    |-- models
    |-- utils
    |-- main.py

Training and Evaluation

If you want to run the model UMKL-G with the loss function power KL divergence in power 2, run the command as below.

python main.py --model UMKL-G --dataset MUTAG --loss_fun PKL --power 2

Results

Check the folder experiemnt, under which results on each dataset are display with

• kernel_list configuration in para.yml
• For each set of parameters, the numbered folder (e.g. 1, 2, ...) contains comparison results of UMKL-G and baselines UMKL, sparse-UMKL.
  • Evaluation metrics include Accuracy, Normalized Mutual Information, and Adjusted Rand Index after clustering on the learned kernel matrix.
  • Losses and weights are recorded as well.

Citation

@inproceedings{sununsupervised,
  title={Unsupervised Multiple Kernel Learning for Graphs via Ordinality Preservation},
  author={Sun, Yan and Kok, Stanley},
  booktitle={The Thirteenth International Conference on Learning Representations}
}