def pointwise_mutual_information(X,
y,
normalize=False,
k_weight=None,
positive=None):
p_c = marginal_estimator(y, smoothing=True)
p_t = marginal_estimator(X, smoothing=True)
p_t.shape = 2, 1
p_c.shape = 1, 2
p_t_c = joint_estimator_point(X, y, smoothing=True)
if k_weight:
p_t_c = p_t_c**k_weight
m = numpy.log(array(p_t_c) / (p_t * p_c))
if normalize:
m = m / -numpy.log(p_t_c)
if positive is 'cutoff':
m[m < .0] = .0
if positive is 'exp':
m = e**m
return array(numpy.max(m, axis=1)).flatten()
def conditional_probabilities(X, y, ratio=False):
p_t_c = joint_estimator_point(X, y, smoothing=True)
p_t = marginal_estimator(X, smoothing=True)
p_t.shape = 2,1
m = p_t_c / p_t
if ratio:
p_c = marginal_estimator(y, smoothing=True)
m = m / p_c
return array(numpy.max(m, axis=1)).flatten()
def conditional_probabilities(X, y, ratio=False):
p_t_c = joint_estimator_point(X, y, smoothing=True)
p_t = marginal_estimator(X, smoothing=True)
p_t.shape = 2, 1
m = p_t_c / p_t
if ratio:
p_c = marginal_estimator(y, smoothing=True)
m = m / p_c
return array(numpy.max(m, axis=1)).flatten()
def mutual_information(X, y):
num_terms = X.shape[1]
num_classes = y.shape[1]
p_c = marginal_estimator(y, smoothing=True)
p_t = marginal_estimator(X, smoothing=True)
p_t_c = joint_estimator_full(X, y, smoothing=True)
ig = zeros((num_terms))
for i in xrange(num_terms):
for j in xrange(num_classes):
ig[i] += p_t_c[0][i, j] * log(p_t_c[0][i, j] / (p_t[i] * p_c[j]))
ig[i] += p_t_c[1][i, j] * log(p_t_c[1][i, j] / (p_t[i] * (1 - p_c[j])))
ig[i] += p_t_c[2][i, j] * log(p_t_c[2][i, j] / ((1 - p_t[i]) * p_c[j]))
ig[i] += p_t_c[3][i, j] * log(p_t_c[3][i, j] / ((1 - p_t[i]) * (1 - p_c[j])))
return ig
def mutual_information(X, y):
num_terms = X.shape[1]
num_classes = y.shape[1]
p_c = marginal_estimator(y, smoothing=True)
p_t = marginal_estimator(X, smoothing=True)
p_t_c = joint_estimator_full(X, y, smoothing=True)
ig = zeros((num_terms))
for i in xrange(num_terms):
for j in xrange(num_classes):
ig[i] += p_t_c[0][i, j] * log(p_t_c[0][i, j] / (p_t[i] * p_c[j]))
ig[i] += p_t_c[1][i, j] * log(p_t_c[1][i, j] / (p_t[i] *
(1 - p_c[j])))
ig[i] += p_t_c[2][i, j] * log(p_t_c[2][i, j] /
((1 - p_t[i]) * p_c[j]))
ig[i] += p_t_c[3][i, j] * log(p_t_c[3][i, j] / ((1 - p_t[i]) *
(1 - p_c[j])))
return ig
def pointwise_mutual_information(X, y, normalize=False, k_weight=None, positive=None):
p_c = marginal_estimator(y, smoothing=True)
p_t = marginal_estimator(X, smoothing=True)
p_t.shape = 2, 1
p_c.shape = 1, 2
p_t_c = joint_estimator_point(X, y, smoothing=True)
if k_weight:
p_t_c = p_t_c**k_weight
m = numpy.log(array(p_t_c) / (p_t * p_c))
if normalize:
m = m / -numpy.log(p_t_c)
if positive is 'cutoff':
m[m < .0] = .0
if positive is 'exp':
m = e**m
return array(numpy.max(m, axis=1)).flatten()
