-
Notifications
You must be signed in to change notification settings - Fork 15
/
kernel_test.py
340 lines (292 loc) · 12.5 KB
/
kernel_test.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
from typing import Optional, Set, Tuple
import numpy as np
import pandas as pd
from scipy import stats
from sklearn.metrics.pairwise import PAIRWISE_KERNEL_FUNCTIONS
from dodiscover.ci.kernel_utils import compute_kernel
from dodiscover.typing import Column
from .base import BaseConditionalIndependenceTest
class KernelCITest(BaseConditionalIndependenceTest):
_allow_multivariate_input: bool = True
def __init__(
self,
kernel_x: str = "rbf",
kernel_y: str = "rbf",
kernel_z: str = "rbf",
null_size: int = 1000,
approx_with_gamma: bool = True,
kwidth_x: Optional[float] = None,
kwidth_y: Optional[float] = None,
kwidth_z: Optional[float] = None,
threshold: float = 1e-5,
n_jobs: Optional[int] = None,
):
"""Kernel (Conditional) Independence Test.
For testing (conditional) independence on continuous data, we
leverage kernels :footcite:`Zhang2011` that are computationally efficient.
Parameters
----------
kernel_x : str, optional
The kernel function for data 'X', by default "rbf".
kernel_y : str, optional
The kernel function for data 'Y', by default "rbf".
kernel_z : str, optional
The kernel function for data 'Z', by default "rbf".
null_size : int, optional
The number of samples to generate for the bootstrap distribution to
approximate the pvalue, by default 1000.
approx_with_gamma : bool, optional
Whether to use the Gamma distribution approximation for the pvalue,
by default True.
kwidth_x : float, optional
The width of the kernel to be applied to the X variable, by default None.
kwidth_y : float, optional
The width of the kernel to be applied to the Y variable, by default None.
kwidth_z : float, optional
The width of the kernel to be applied to the Z variable, by default None.
threshold : float, optional
The threshold set on the value of eigenvalues, by default 1e-5. Used
to regularize the method.
n_jobs : int, optional
The number of CPUs to use, by default None.
Notes
-----
Valid strings for ``compute_kernel`` are, as defined in
:func:`sklearn.metrics.pairwise.pairwise_kernels`,
[``"additive_chi2"``, ``"chi2"``, ``"linear"``, ``"poly"``,
``"polynomial"``, ``"rbf"``,
``"laplacian"``, ``"sigmoid"``, ``"cosine"``]
References
----------
.. footbibliography::
"""
if isinstance(kernel_x, str) and kernel_x not in PAIRWISE_KERNEL_FUNCTIONS:
raise ValueError(
f"The kernels that are currently supported are {PAIRWISE_KERNEL_FUNCTIONS}. "
f"You passed in {kernel_x} for kernel_x."
)
if isinstance(kernel_y, str) and kernel_y not in PAIRWISE_KERNEL_FUNCTIONS:
raise ValueError(
f"The kernels that are currently supported are {PAIRWISE_KERNEL_FUNCTIONS}. "
f"You passed in {kernel_y} for kernel_y."
)
if isinstance(kernel_z, str) and kernel_z not in PAIRWISE_KERNEL_FUNCTIONS:
raise ValueError(
f"The kernels that are currently supported are {PAIRWISE_KERNEL_FUNCTIONS}. "
f"You passed in {kernel_z} for kernel_z."
)
self.kernel_x = kernel_x
self.kernel_y = kernel_y
self.kernel_z = kernel_z
self.null_size = null_size
self.approx_with_gamma = approx_with_gamma
self.threshold = threshold
self.n_jobs = n_jobs
# hyperparameters of the kernsl
self.kwidth_x = kwidth_x
self.kwidth_y = kwidth_y
self.kwidth_z = kwidth_z
def test(
self,
df: pd.DataFrame,
x_vars: Set[Column],
y_vars: Set[Column],
z_covariates: Optional[Set[Column]] = None,
) -> Tuple[float, float]:
"""Run CI test.
Parameters
----------
df : pd.DataFrame
The dataframe containing the dataset.
x_vars : Set of column
A column in ``df``.
y_vars : Set of column
A column in ``df``.
z_covariates : Set, optional
A set of columns in ``df``, by default None. If None, then
the test should run a standard independence test.
Returns
-------
stat : float
The test statistic.
pvalue : float
The p-value of the test.
"""
self._check_test_input(df, x_vars, y_vars, z_covariates)
if z_covariates is None or len(z_covariates) == 0:
Z = None
else:
Z = df[list(z_covariates)].to_numpy().reshape((-1, len(z_covariates)))
x_columns = list(x_vars)
y_columns = list(y_vars)
X = df[x_columns].to_numpy()
Y = df[y_columns].to_numpy()
if X.ndim == 1:
X = X[:, np.newaxis]
if Y.ndim == 1:
Y = Y[:, np.newaxis]
# first normalize the data to have zero mean and unit variance
# along the columns of the data
X = stats.zscore(X, axis=0)
Y = stats.zscore(Y, axis=0)
if Z is not None:
Z = stats.zscore(Z, axis=0)
# when running CI, \ddot{X} comprises of (X, Z)
X = np.concatenate((X, Z), axis=1)
Kz, sigma_z = compute_kernel(
Z,
distance_metric="l2",
metric=self.kernel_z,
kwidth=self.kwidth_z,
centered=True,
n_jobs=self.n_jobs,
)
# compute the centralized kernel matrices of each the datasets
Kx, sigma_x = compute_kernel(
X,
distance_metric="l2",
metric=self.kernel_x,
kwidth=self.kwidth_x,
centered=True,
n_jobs=self.n_jobs,
)
Ky, sigma_y = compute_kernel(
Y,
distance_metric="l2",
metric=self.kernel_y,
kwidth=self.kwidth_y,
centered=True,
n_jobs=self.n_jobs,
)
if Z is None:
# test statistic is just the normal bivariate independence
# test statistic
test_stat = self._compute_V_statistic(Kx, Ky)
if self.approx_with_gamma:
# approximate the pvalue using the Gamma distribution
k_appr, theta_appr = self._approx_gamma_params_ind(Kx, Ky)
pvalue = 1 - stats.gamma.cdf(test_stat, k_appr, 0, theta_appr)
else:
null_samples = self._compute_null_ind(Kx, Ky, n_samples=self.null_size)
pvalue = np.sum(null_samples > test_stat) / float(self.null_size)
else:
# compute the centralizing matrix for the kernels according to
# conditioning set Z
epsilon = 1e-6
n = Kx.shape[0]
Rz = epsilon * np.linalg.pinv(Kz + epsilon * np.eye(n))
# compute the centralized kernel matrices
KxzR = Rz.dot(Kx).dot(Rz)
KyzR = Rz.dot(Ky).dot(Rz)
# compute the conditional independence test statistic
test_stat = self._compute_V_statistic(KxzR, KyzR)
# compute the product of the eigenvectors
uu_prod = self._compute_prod_eigvecs(KxzR, KyzR, threshold=self.threshold)
if self.approx_with_gamma:
# approximate the pvalue using the Gamma distribution
k_appr, theta_appr = self._approx_gamma_params_ci(uu_prod)
pvalue = 1 - stats.gamma.cdf(test_stat, k_appr, 0, theta_appr)
else:
null_samples = self._compute_null_ci(uu_prod, self.null_size)
pvalue = np.sum(null_samples > test_stat) / float(self.null_size)
return test_stat, pvalue
def _approx_gamma_params_ind(self, Kx, Ky):
T = Kx.shape[0]
mean_appr = np.trace(Kx) * np.trace(Ky) / T
var_appr = 2 * np.trace(Kx.dot(Kx)) * np.trace(Ky.dot(Ky)) / T / T
k_appr = mean_appr**2 / var_appr
theta_appr = var_appr / mean_appr
return k_appr, theta_appr
def _approx_gamma_params_ci(self, uu_prod):
"""Get parameters of the approximated Gamma distribution.
Parameters
----------
uu_prod : np.ndarray of shape (n_features, n_features)
The product of the eigenvectors of Kx and Ky, the kernels
on the input data, X and Y.
Returns
-------
k_appr : float
The shape parameter of the Gamma distribution.
theta_appr : float
The scale parameter of the Gamma distribution.
Notes
-----
X ~ Gamma(k, theta) with a probability density function of the following:
.. math::
f(x; k, \\theta) = \\frac{x^{k-1} e^{-x / \\theta}}{\\theta^k \\Gamma(k)}
where $\\Gamma(k)$ is the Gamma function evaluated at k. In this scenario
k governs the shape of the pdf, while $\\theta$ governs more how spread out
the data is.
"""
# approximate the mean and the variance
mean_appr = np.trace(uu_prod)
var_appr = 2 * np.trace(uu_prod.dot(uu_prod))
k_appr = mean_appr**2 / var_appr
theta_appr = var_appr / mean_appr
return k_appr, theta_appr
def _compute_prod_eigvecs(self, Kx, Ky, threshold=None):
T = Kx.shape[0]
wx, vx = np.linalg.eigh(0.5 * (Kx + Kx.T))
wy, vy = np.linalg.eigh(0.5 * (Ky + Ky.T))
if threshold is not None:
# threshold eigenvalues that are below a certain threshold
# and remove their corresponding values and eigenvectors
vx = vx[:, wx > np.max(wx) * threshold]
wx = wx[wx > np.max(wx) * threshold]
vy = vy[:, wy > np.max(wy) * threshold]
wy = wy[wy > np.max(wy) * threshold]
# scale the eigenvectors by their eigenvalues
vx = vx.dot(np.diag(np.sqrt(wx)))
vy = vy.dot(np.diag(np.sqrt(wy)))
# compute the product of the scaled eigenvectors
num_eigx = vx.shape[1]
num_eigy = vy.shape[1]
size_u = num_eigx * num_eigy
uu = np.zeros((T, size_u))
for i in range(0, num_eigx):
for j in range(0, num_eigy):
# compute the dot product of eigenvectors
uu[:, i * num_eigy + j] = vx[:, i] * vy[:, j]
# now compute the product
if size_u > T:
uu_prod = uu.dot(uu.T)
else:
uu_prod = uu.T.dot(uu)
return uu_prod
def _compute_V_statistic(self, KxR, KyR):
# n = KxR.shape[0]
# compute the sum of the two kernsl
Vstat = np.sum(KxR * KyR)
return Vstat
def _compute_null_ind(self, Kx, Ky, n_samples, max_num_eigs=1000):
n = Kx.shape[0]
# get the eigenvalues in ascending order, smallest to largest
eigvals_x = np.linalg.eigvalsh(Kx)
eigvals_y = np.linalg.eigvalsh(Ky)
# optionally only keep the largest "N" eigenvalues
eigvals_x = eigvals_x[-max_num_eigs:]
eigvals_y = eigvals_y[-max_num_eigs:]
num_eigs = len(eigvals_x)
# compute the entry-wise product of the eigenvalues and store it as a vector
eigvals_prod = np.dot(
eigvals_x.reshape(num_eigs, 1), eigvals_y.reshape(1, num_eigs)
).reshape((-1, 1))
# only keep eigenvalues above a certain threshold
eigvals_prod = eigvals_prod[eigvals_prod > eigvals_prod.max() * self.threshold]
# generate chi-square distributed values z_{ij} with degree of freedom 1
f_rand = np.random.chisquare(df=1, size=(len(eigvals_prod), n_samples))
# compute the null distribution consisting now of (n_samples)
# of chi-squared random variables weighted by the eigenvalue products
null_dist = 1.0 / n * eigvals_prod.T.dot(f_rand)
return null_dist
def _compute_null_ci(self, uu_prod, n_samples):
# the eigenvalues of ww^T
eig_uu = np.linalg.eigvalsh(uu_prod)
eig_uu = eig_uu[eig_uu > eig_uu.max() * self.threshold]
# generate chi-square distributed values z_{ij} with degree of freedom 1
f_rand = np.random.chisquare(df=1, size=(eig_uu.shape[0], n_samples))
# compute the null distribution consisting now of (n_samples)
# of chi-squared random variables weighted by the eigenvalue products
null_dist = eig_uu.T.dot(f_rand)
return null_dist