scikit-learn/examples/svm/plot_svm_scale_c.py

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

169 lines
5.3 KiB
Python
Raw Normal View History

r"""
2013-01-15 21:36:00 +08:00
==============================================
Scaling the regularization parameter for SVCs
==============================================
2012-07-25 18:50:00 +08:00
The following example illustrates the effect of scaling the
2012-07-25 22:43:42 +08:00
regularization parameter when using :ref:`svm` for
:ref:`classification <svm_classification>`.
2012-07-25 18:50:00 +08:00
For SVC classification, we are interested in a risk minimization for the
equation:
.. math::
C \sum_{i=1, n} \mathcal{L} (f(x_i), y_i) + \Omega (w)
where
- :math:`C` is used to set the amount of regularization
- :math:`\mathcal{L}` is a `loss` function of our samples
and our model parameters.
- :math:`\Omega` is a `penalty` function of our model parameters
2012-07-25 22:43:42 +08:00
If we consider the loss function to be the individual error per
sample, then the data-fit term, or the sum of the error for each sample, will
increase as we add more samples. The penalization term, however, will not
2012-07-25 18:50:00 +08:00
increase.
When using, for example, :ref:`cross validation <cross_validation>`, to
2012-09-04 20:00:41 +08:00
set the amount of regularization with `C`, there will be a
2012-08-27 23:33:03 +08:00
different amount of samples between the main problem and the smaller problems
within the folds of the cross validation.
2012-07-25 18:50:00 +08:00
Since our loss function is dependent on the amount of samples, the latter
2012-07-25 22:43:42 +08:00
will influence the selected value of `C`.
2012-07-25 18:50:00 +08:00
The question that arises is `How do we optimally adjust C to
2012-09-04 20:00:41 +08:00
account for the different amount of training samples?`
2012-07-25 18:50:00 +08:00
The figures below are used to illustrate the effect of scaling our
2012-08-09 20:48:17 +08:00
`C` to compensate for the change in the number of samples, in the
case of using an `l1` penalty, as well as the `l2` penalty.
2012-07-25 18:50:00 +08:00
l1-penalty case
2012-07-25 18:50:00 +08:00
-----------------
In the `l1` case, theory says that prediction consistency
2012-07-25 18:50:00 +08:00
(i.e. that under given hypothesis, the estimator
2012-09-04 20:00:41 +08:00
learned predicts as well as a model knowing the true distribution)
is not possible because of the bias of the `l1`. It does say, however,
2012-08-09 20:48:17 +08:00
that model consistency, in terms of finding the right set of non-zero
2012-07-25 22:43:42 +08:00
parameters as well as their signs, can be achieved by scaling
`C1`.
2012-07-25 18:50:00 +08:00
l2-penalty case
2012-07-25 18:50:00 +08:00
-----------------
2012-09-04 20:00:41 +08:00
The theory says that in order to achieve prediction consistency, the
penalty parameter should be kept constant
as the number of samples grow.
2012-07-25 18:50:00 +08:00
Simulations
------------
2012-07-25 22:43:42 +08:00
The two figures below plot the values of `C` on the `x-axis` and the
2012-07-25 18:50:00 +08:00
corresponding cross-validation scores on the `y-axis`, for several different
fractions of a generated data-set.
In the `l1` penalty case, the cross-validation-error correlates best with
2012-08-27 23:33:03 +08:00
the test-error, when scaling our `C` with the number of samples, `n`,
which can be seen in the first figure.
2012-07-25 18:50:00 +08:00
For the `l2` penalty case, the best result comes from the case where `C`
2012-07-25 18:50:00 +08:00
is not scaled.
2012-07-25 22:43:42 +08:00
.. topic:: Note:
2012-07-25 18:50:00 +08:00
Two separate datasets are used for the two different plots. The reason
behind this is the `l1` case works better on sparse data, while `l2`
2012-07-25 22:43:42 +08:00
is better suited to the non-sparse case.
2012-07-25 18:50:00 +08:00
"""
2012-07-25 18:50:00 +08:00
# Author: Andreas Mueller <amueller@ais.uni-bonn.de>
# Jaques Grobler <jaques.grobler@inria.fr>
# License: BSD 3 clause
2012-07-25 18:50:00 +08:00
import numpy as np
import matplotlib.pyplot as plt
2012-07-25 18:50:00 +08:00
from sklearn.svm import LinearSVC
from sklearn.model_selection import ShuffleSplit
from sklearn.model_selection import GridSearchCV
2012-07-25 18:50:00 +08:00
from sklearn.utils import check_random_state
from sklearn import datasets
rnd = check_random_state(1)
# set up dataset
n_samples = 100
2012-07-25 21:18:45 +08:00
n_features = 300
2012-07-25 22:43:42 +08:00
# l1 data (only 5 informative features)
2012-09-05 01:39:25 +08:00
X_1, y_1 = datasets.make_classification(
2012-12-25 20:16:05 +08:00
n_samples=n_samples, n_features=n_features, n_informative=5, random_state=1
)
2012-07-25 22:43:42 +08:00
# l2 data: non sparse, but less features
2012-07-25 21:18:45 +08:00
y_2 = np.sign(0.5 - rnd.rand(n_samples))
X_2 = rnd.randn(n_samples, n_features // 5) + y_2[:, np.newaxis]
X_2 += 5 * rnd.randn(n_samples, n_features // 5)
2012-07-25 22:43:42 +08:00
clf_sets = [
(
LinearSVC(penalty="l1", loss="squared_hinge", dual=False, tol=1e-3),
2012-09-04 20:34:42 +08:00
np.logspace(-2.3, -1.3, 10),
X_1,
y_1,
),
(
LinearSVC(penalty="l2", loss="squared_hinge", dual=True),
2012-07-25 21:18:45 +08:00
np.logspace(-4.5, -2, 10),
X_2,
y_2,
),
2012-07-25 21:18:45 +08:00
]
2015-10-22 19:59:52 +08:00
colors = ["navy", "cyan", "darkorange"]
lw = 2
2012-07-25 18:50:00 +08:00
for clf, cs, X, y in clf_sets:
2012-07-25 18:50:00 +08:00
# set up the plot for each regressor
fig, axes = plt.subplots(nrows=2, sharey=True, figsize=(9, 10))
2012-07-25 22:43:42 +08:00
2012-07-25 21:18:45 +08:00
for k, train_size in enumerate(np.linspace(0.3, 0.7, 3)[::-1]):
2012-07-25 18:50:00 +08:00
param_grid = dict(C=cs)
2012-07-25 21:18:45 +08:00
# To get nice curve, we need a large number of iterations to
# reduce the variance
2012-07-25 18:50:00 +08:00
grid = GridSearchCV(
clf,
refit=False,
param_grid=param_grid,
cv=ShuffleSplit(
train_size=train_size, test_size=0.3, n_splits=50, random_state=1
),
)
2012-07-25 18:50:00 +08:00
grid.fit(X, y)
scores = grid.cv_results_["mean_test_score"]
2012-07-25 22:43:42 +08:00
scales = [
(1, "No scaling"),
((n_samples * train_size), "1/n_samples"),
2012-07-25 18:50:00 +08:00
]
for ax, (scaler, name) in zip(axes, scales):
ax.set_xlabel("C")
ax.set_ylabel("CV Score")
2012-09-05 01:39:25 +08:00
grid_cs = cs * float(scaler) # scale the C's
ax.semilogx(
grid_cs,
scores,
label="fraction %.2f" % train_size,
color=colors[k],
lw=lw,
)
ax.set_title(
"scaling=%s, penalty=%s, loss=%s" % (name, clf.penalty, clf.loss)
)
2012-07-25 18:50:00 +08:00
plt.legend(loc="best")
plt.show()