minirank/minirank/datasets.py at master · fabianp/minirank

Repository navigation

import numpy as np

def sigmoid_with_noise(n_samples, n_features, outliers=0.,

seed=None, noise_amplitude=.2, slope=1.):

"""

outliers: floating-point in [0, 1]

fraction of outliers

"""

np.random.seed(seed)

w = np.random.rand(n_features)

#w /= np.linalg.norm(w)

y = 2 * (np.random.rand(n_samples) - 0.5)

p = y[:, np.newaxis] * w

# centering at [0, 1]

p -= np.min(p)

p /= np.max(p)

#p /= 100

# p lies in [0, 1] but we'd like it to lie in [epsilon, 1 - epsilon]

# with epsilon a small quantity

epsilon = 1e-6 / 2

m = 1 / (1 + epsilon)

p *= m

p += (1 - m) / 2

#assert np.abs(p.min() - (1 - p.max())) < 1e-3

X = np.log(p) - np.log(1 - p) + noise_amplitude * np.random.rand(n_samples, n_features)

# normalize features

# for i in range(X.shape[1]):

# X[:, i] /= np.linalg.norm(X[:, i])

for _ in range(int(outliers * n_samples)):

a = np.random.randint(0, n_samples)

b = np.random.randint(0, n_samples)

tmp = y[b]

y[b] = y[a]

y[a] = tmp

return X, y, w