Skip to content

Latest commit

 

History

History
49 lines (33 loc) · 2.67 KB

README.md

File metadata and controls

49 lines (33 loc) · 2.67 KB
Raw

CARL-G: Clustering-Accelerated Representation Learning on Graphs

CARL-G Diagram

Implementation Details

As described in the paper, we run all of our non-timing experiments on a comination of local and cloud resources - either on a NVIDIA RTX A4000 or V100 GPU (each with 16GB of VRAM). All timing experiments were conducted on Google Cloud Platform with a V100 GPU. We perform 50 runs of Bayesian hyperparameter optimization using Weights and Biases on each CVI/dataset/task for each of the three methods. The best-performing hyperparameters for each CVI and dataset on node classification can be found in the config directory.

Running the Code

The code for CARL-G is in the src directory. train_carlg.py contains the code for the base CARL-G implementation, which trains and evaluates a model for node classification. train_carlg_clustering.py trains and evaluates CARL-G on node clustering and similarity search.

To run the model with the best hyperparameters, you can use the --flagfile flag. For example:

cd src
python train_carlg.py --flagfile=../config/sim_amazon-computers_class.cfg

to train and evaluate the model on the amazon-computers dataset for node classification tasks.

Many configurations of the models can be modified through the use of flags:

  • --clustering_metric: This specifies the variant of CARL-G. The three possible values are:
    • silhouette: $\text{CARL-G}_{Sil}$ - the silhouette-based variant.
    • fast_silhouette: $\text{CARL-G}_{Sim}$- the faster simplified-silhouette-based variant.
    • vrc: $\text{CARL-G}_{VRC}$ - the VRC-based variant.
  • --score_target: The target CVI score ($\tau$ in the paper).
  • --num_clusters: The number of clusters ($c$ in the paper).
  • --dataset: Which dataset to run the model on.

There are more configuration options that can be seen by viewing the code or by using the --helpfull flag.

We also include the weights for classification tasks in the weights directory, which can be directly loaded for evaluation purposes.

Credit

We used code from Link Prediction with Non-Contrastive Learning. We also used the evaluation portions of the code from the reference BGRL implementation and the reference AFGRL implementation.

If you use this code, please cite:

@article{shiao2023carl,
  title={CARL-G: Clustering-Accelerated Representation Learning on Graphs},
  author={Shiao, William and Saini, Uday Singh and Liu, Yozen and Zhao, Tong and Shah, Neil and Papalexakis, Evangelos E},
  journal={arXiv preprint arXiv:2306.06936},
  year={2023}
}