Active Graph Learning

Pytorch implementation of the paper "MetAL: Active Semi-Supervised Learning on Graphs via Meta Learning" (2020)

Dependencies

1. python 3.6+
2. pytorch 1.4+
3. numpy
4. scipy
5. networkx
6. scikit-learn
7. timebudget

Run the code

Make sure you have installed all requirements.

Run an example experiment with: sh run_active_learn.sh

This will run 2 trials of meta active learning on CiteSeer dataset and save the performance to a csv file.

Output

Once you execute an active learning experiment, performance is saved as a csv file in the results directory. For example, running 10 trials of entropy acquisition function will create a csv file with the name citeseer_entropy_10trials-accuracy.csv.

Structure of the results file

Each row corresponds to an acquisition of a set of nodes. For each acquisition accuracy, macro-f1, and micro-f1 of the test set is saved along columns of the csv file.

Visualization of results

The Jupyter notebook notebooks/analysis/performance_summary.ipynb is used to load results CSV files and to plot how the performance (accuracy and macro-f1) varies with acquisition of labels of the unlabeled nodes.

Cite

Please cite our paper if you use this code in your own work:

 @InProceedings{madhawa20metal,
title = {{M}et{A}{L}: {A}ctive {S}emi-{S}upervised {L}earning on {G}raphs via {M}eta-{L}earning},
author = {Madhawa, Kaushalya and Murata, Tsuyoshi}, 
booktitle = {Proceedings of The 12th Asian Conference on Machine Learning}, 
pages = {561--576}, 
year = {2020}, 
editor = {Sinno Jialin Pan and Masashi Sugiyama}, 
volume = {129}, 
series = {Proceedings of Machine Learning Research}, 
address = {Bangkok, Thailand}, month = {18--20 Nov}, 
publisher = {PMLR}, 
pdf = {http://proceedings.mlr.press/v129/madhawa20a/madhawa20a.pdf}, 
url = {http://proceedings.mlr.press/v129/madhawa20a.html}}