Python code for doing active learning in graph-based semi-supervised learning (GBSSL) paradigm. Implements testing done in paper that will soon appear and be submitted for peer-review.

To run tests in this framework, run scripts bin_run.py or multi_run.py specifying location (--data-root) of .npz file that contains variables X (N x d numpy array) and labels (N vector numpy array). Default is hard-coded in the scripts to run on all possible acquisition functions, but can change the list variable acq_models

• possible choices for acq_models:
  • acquisitions functions : mc (Model Change), uncertainty (Uncertainty), vopt (VOpt), sopt (SigmaOpt), rand (Random)
  • binary models : gr (Gaussian Regression), log (Logistic Loss), probitnorm (Probit - Normal)
  • multiclass models : gr(Gaussian Regression), ce (Cross-Entropy)
  • Separate acquisition function and model with double-dash: e.g. mc--gr --> Model Change acquisition function in Gaussian Regression Model.

This repo requires the Python packages: sklearn, mlflow, numpy, scipy.

Can simply run test on the synthetic dataset "Binary Clusters" presented in the paper:

python binary_clusters_run.py --al-iters 100 --B 1 # run sequential active learning on binary clusters data for 100 active learning iterations
python binary_clusters_run.py --al-iters 20 --B 5 # run batch activate learning on binary clusters data for 20 active learning iterations

Example plots from code in results/acc_figures.py (which can be opened as a Jupyter notebook)

MNIST	Salinas A	Urban

MNIST	Salinas A	Urban