GEEM : Active Learning for Graphs

Some nodes are more informative than others.

Semi-supervised node classification is the problem of infering node labels $\mathbf{y}$ over the whole graph given the graph structured $\mathcal{G}$, node attributes $\mathbf{X}$, and the labels of some of the nodes $\mathbf{y}_{\mathcal{L}}$.

Carefully selecting which labeled nodes are part of the dataset $Y_{\mathcal{L}}$ with active learning can greatly reduce the required number of data to reach the same accuracy.

Our GEEM algorithm uses graph cognizant logistic regression, equivalent to a linearized graph convolutional neural network (GCN), for the prediction phase and maximizes the expected error reduction in the query phase.

You can view our GEEM algorithm in action against a random baseline:

Authors:

• Florence Regol
• Soumyasundar Pal
• Yingxue Zhang
• Mark Coates (Prof. McGill University)

Citation

This project was published at ICML 2020.

@inproceedings{regol2020geem,
  title = {Active Learning on Attributed Graphs via Graph Cognizant Logistic Regression and Preemptive Query Generation},
  author = {Regol, Florence and Pal, Soumyasundar and Zhang, Yingxue and Coates, Mark},
  booktitle = {International Conference on Machine Learning (ICML)},
  pages = {8041--8050},
  year = {2020},
  month = {Jul}
}

ArXiv link