gaussianfield

Python implementation of the Gaussian field classifier from Zhu 2003.

Zhu, X; Lafferty, JK; Ghahramani, Z; (2003) Combining Active Learning and Semi-Supervised Learning Using Gaussian Fields and Harmonic Functions. In: (Proceedings) 20th International Conference on Machine Learning workshop.

http://www.cs.cmu.edu/~zhuxj/pub/semisupervisedcode/active_learning/

Brief usage snippet

Main inputs are a kernel matrix K for the full dataset, one-hot-encoded class labels for the observed examples, and an indicator vector for the observed examples. Unlike the Matlab implementation, observed and unobserved data don't need to be contiguous.

import gaussianfield
from keras.utils.np_utils import to_categorical

# load data, labels, compute kernel matrix...
X, labels = load_dataset(...)
N, ndim = X.shape
K = compute_kernel_matrix(X)

# choose some random examples for initial observed set...
thresh = 0.05
observed = np.random.random(N) < thresh
N_observed = observed.sum()

field_solution, inverse_laplacian = gaussianfield.solve(K, labels[observed], observed)
class_predictions = np.argmax(field, axis=1)

# compute expected risk for active learning...
risk = gaussianfield.expected_risk(field_solution, inverse_laplacian)

# get the query index relative to the full dataset
_risk = 1000 * np.ones(labels.size)
_risk[~observed] = risk
query_idx = np.argmin(_risk)