# Natural Language Toolkit: Expectation Maximization Clusterer # # Copyright (C) 2001-2010 NLTK Project # Author: Trevor Cohn # URL: # For license information, see LICENSE.TXT import numpy from api import * from util import * class EMClusterer(VectorSpaceClusterer): """ The Gaussian EM clusterer models the vectors as being produced by a mixture of k Gaussian sources. The parameters of these sources (prior probability, mean and covariance matrix) are then found to maximise the likelihood of the given data. This is done with the expectation maximisation algorithm. It starts with k arbitrarily chosen means, priors and covariance matrices. It then calculates the membership probabilities for each vector in each of the clusters; this is the 'E' step. The cluster parameters are then updated in the 'M' step using the maximum likelihood estimate from the cluster membership probabilities. This process continues until the likelihood of the data does not significantly increase. """ def __init__(self, initial_means, priors=None, covariance_matrices=None, conv_threshold=1e-6, bias=0.1, normalise=False, svd_dimensions=None): """ Creates an EM clusterer with the given starting parameters, convergence threshold and vector mangling parameters. @param initial_means: the means of the gaussian cluster centers @type initial_means: [seq of] numpy array or seq of SparseArray @param priors: the prior probability for each cluster @type priors: numpy array or seq of float @param covariance_matrices: the covariance matrix for each cluster @type covariance_matrices: [seq of] numpy array @param conv_threshold: maximum change in likelihood before deemed convergent @type conv_threshold: int or float @param bias: variance bias used to ensure non-singular covariance matrices @type bias: float @param normalise: should vectors be normalised to length 1 @type normalise: boolean @param svd_dimensions: number of dimensions to use in reducing vector dimensionsionality with SVD @type svd_dimensions: int """ VectorSpaceClusterer.__init__(self, normalise, svd_dimensions) self._means = numpy.array(initial_means, numpy.float64) self._num_clusters = len(initial_means) self._conv_threshold = conv_threshold self._covariance_matrices = covariance_matrices self._priors = priors self._bias = bias def num_clusters(self): return self._num_clusters def cluster_vectorspace(self, vectors, trace=False): assert len(vectors) > 0 # set the parameters to initial values dimensions = len(vectors[0]) means = self._means priors = self._priors if not priors: priors = self._priors = numpy.ones(self._num_clusters, numpy.float64) / self._num_clusters covariances = self._covariance_matrices if not covariances: covariances = self._covariance_matrices = \ [ numpy.identity(dimensions, numpy.float64) for i in range(self._num_clusters) ] # do the E and M steps until the likelihood plateaus lastl = self._loglikelihood(vectors, priors, means, covariances) converged = False while not converged: if trace: print 'iteration; loglikelihood', lastl # E-step, calculate hidden variables, h[i,j] h = numpy.zeros((len(vectors), self._num_clusters), numpy.float64) for i in range(len(vectors)): for j in range(self._num_clusters): h[i,j] = priors[j] * self._gaussian(means[j], covariances[j], vectors[i]) h[i,:] /= sum(h[i,:]) # M-step, update parameters - cvm, p, mean for j in range(self._num_clusters): covariance_before = covariances[j] new_covariance = numpy.zeros((dimensions, dimensions), numpy.float64) new_mean = numpy.zeros(dimensions, numpy.float64) sum_hj = 0.0 for i in range(len(vectors)): delta = vectors[i] - means[j] new_covariance += h[i,j] * \ numpy.multiply.outer(delta, delta) sum_hj += h[i,j] new_mean += h[i,j] * vectors[i] covariances[j] = new_covariance / sum_hj means[j] = new_mean / sum_hj priors[j] = sum_hj / len(vectors) # bias term to stop covariance matrix being singular covariances[j] += self._bias * \ numpy.identity(dimensions, numpy.float64) # calculate likelihood - FIXME: may be broken l = self._loglikelihood(vectors, priors, means, covariances) # check for convergence if abs(lastl - l) < self._conv_threshold: converged = True lastl = l def classify_vectorspace(self, vector): best = None for j in range(self._num_clusters): p = self._priors[j] * self._gaussian(self._means[j], self._covariance_matrices[j], vector) if not best or p > best[0]: best = (p, j) return best[1] def likelihood_vectorspace(self, vector, cluster): cid = self.cluster_names().index(cluster) return self._priors[cluster] * self._gaussian(self._means[cluster], self._covariance_matrices[cluster], vector) def _gaussian(self, mean, cvm, x): m = len(mean) assert cvm.shape == (m, m), \ 'bad sized covariance matrix, %s' % str(cvm.shape) try: det = numpy.linalg.det(cvm) inv = numpy.linalg.inv(cvm) a = det ** -0.5 * (2 * numpy.pi) ** (-m / 2.0) dx = x - mean print dx, inv b = -0.5 * numpy.dot( numpy.dot(dx, inv), dx) return a * numpy.exp(b) except OverflowError: # happens when the exponent is negative infinity - i.e. b = 0 # i.e. the inverse of cvm is huge (cvm is almost zero) return 0 def _loglikelihood(self, vectors, priors, means, covariances): llh = 0.0 for vector in vectors: p = 0 for j in range(len(priors)): p += priors[j] * \ self._gaussian(means[j], covariances[j], vector) llh += numpy.log(p) return llh def __repr__(self): return '' % list(self._means) def demo(): """ Non-interactive demonstration of the clusterers with simple 2-D data. """ from nltk import cluster # example from figure 14.10, page 519, Manning and Schutze vectors = [numpy.array(f) for f in [[0.5, 0.5], [1.5, 0.5], [1, 3]]] means = [[4, 2], [4, 2.01]] clusterer = cluster.EMClusterer(means, bias=0.1) clusters = clusterer.cluster(vectors, True, trace=True) print 'Clustered:', vectors print 'As: ', clusters print for c in range(2): print 'Cluster:', c print 'Prior: ', clusterer._priors[c] print 'Mean: ', clusterer._means[c] print 'Covar: ', clusterer._covariance_matrices[c] print # classify a new vector vector = numpy.array([2, 2]) print 'classify(%s):' % vector, print clusterer.classify(vector) # show the classification probabilities vector = numpy.array([2, 2]) print 'classification_probdist(%s):' % vector pdist = clusterer.classification_probdist(vector) for sample in pdist.samples(): print '%s => %.0f%%' % (sample, pdist.prob(sample) *100) # # The following demo code is broken. # # # use a set of tokens with 2D indices # vectors = [numpy.array(f) for f in [[3, 3], [1, 2], [4, 2], [4, 0], [2, 3], [3, 1]]] # # test the EM clusterer with means given by k-means (2) and # # dimensionality reduction # clusterer = cluster.KMeans(2, euclidean_distance, svd_dimensions=1) # print 'Clusterer:', clusterer # clusters = clusterer.cluster(vectors) # means = clusterer.means() # print 'Means:', clusterer.means() # print # clusterer = cluster.EMClusterer(means, svd_dimensions=1) # clusters = clusterer.cluster(vectors, True) # print 'Clusterer:', clusterer # print 'Clustered:', str(vectors)[:60], '...' # print 'As:', str(clusters)[:60], '...' # print # # classify a new vector # vector = numpy.array([3, 3]) # print 'classify(%s):' % vector, # print clusterer.classify(vector) # print # # show the classification probabilities # vector = numpy.array([2.2, 2]) # print 'classification_probdist(%s)' % vector # pdist = clusterer.classification_probdist(vector) # for sample in pdist: # print '%s => %.0f%%' % (sample, pdist.prob(sample) *100) if __name__ == '__main__': demo()