mirror of
https://asciireactor.com/otho/cs-5821.git
synced 2025-10-31 17:48:04 +00:00
140 lines
6.9 KiB
Plaintext
140 lines
6.9 KiB
Plaintext
We now review k-fold cross-validation.
|
|
|
|
──────────────────────────────────────────────────────────────────────────
|
|
(a) Explain how k-fold cross-validation is implemented.
|
|
|
|
The data are split into k groups of about equal size. The
|
|
first group is set aside to be used as a validation set. The
|
|
model is trained on the remaining sets, and an MSE is
|
|
computed using the validation set. This process is repeated
|
|
using the next set as a validation set, producing a second
|
|
MSE. The process is performed on each of the k groups the
|
|
data were split into, producing k MSEs, one for each
|
|
validation set. Given a good value of k, this mean squared
|
|
error is a good approximation for the minimum point along
|
|
the flexibility axis that allows us to judge the appropriate
|
|
flexibility to model some data, i.e., to estimate the right
|
|
bias-variance tradeoff for a particular model.
|
|
|
|
──────────────────────────────────────────────────────────────────────────
|
|
(b) What are the advantages and disadvantages of k-fold cross-
|
|
validation relative to:
|
|
|
|
i. The validation set approach?
|
|
|
|
The validation set approach is basically k-fold with k=2.
|
|
This gives a biased measure of the test error rate. However,
|
|
it's very quick to compute. Of course, the MSE here will
|
|
have no variance because we are only taking a single MSE
|
|
from a single test set: so this gives large bias of the MSE
|
|
but no variance.
|
|
|
|
ii. LOOCV?
|
|
|
|
LOOCV is on the other end of the spectrum from this. It
|
|
gives a very unbiased MSE for each validation set because
|
|
each validation set is a single datum and is being tested
|
|
against a nearly-full training set, compared to the overall
|
|
data set. However, the MSEs will be highly-correlated
|
|
because the training sets are almost completely the same,
|
|
and this guarantees a high variance.
|
|
|
|
K-fold cross-validation occupies the space in-between, where
|
|
it has some bias and some variance of the MSE. Empirically,
|
|
we know that k=5 and k=10 are two values that often work
|
|
well to provide good estimates of these MSEs, because they
|
|
provide balanced bias and variance.
|
|
|
|
|
|
|
|
|
|
6. We continue to consider the use of a logistic regression model to
|
|
predict the probability of default using income and balance on
|
|
the Default data set. In particular, we will now compute
|
|
estimates for the standard errors of the income and balance
|
|
logistic regression co- efficients in two different ways: (1)
|
|
using the bootstrap, and (2) using the standard formula for
|
|
computing the standard errors in the glm() function. Do not
|
|
forget to set a random seed before beginning your analysis.
|
|
|
|
──────────────────────────────────────────────────────────────────────────
|
|
(a) Using the summary() and glm() functions, determine the esti-
|
|
mated standard errors for the coefficients associated with
|
|
income and balance in a multiple logistic regression model that
|
|
uses both predictors.
|
|
|
|
|
|
10,000 entries in the table, so testing 2 ways to separate the data,
|
|
|
|
> train.default1 = Default[1:6001,]
|
|
> train.default2 = Default[1:5001,]
|
|
> test.default1 = Default[6002:10000,]
|
|
> test.default2 = Default[5002:10000,]
|
|
|
|
> fit.glm.default1 = glm(default ~ income + balance,train.default1, family="binomial")
|
|
> fit.glm.default2 = glm(default ~ income + balance,train.default2, family="binomial")
|
|
|
|
*** from summary(fit.glm.default1), we get std. error
|
|
|
|
Coefficients:
|
|
Estimate Std. Error
|
|
(Intercept) -1.156e+01 5.622e-01
|
|
income 2.237e-05 6.491e-06
|
|
balance 5.649e-03 2.916e-04
|
|
|
|
|
|
*** from summary(fit.glm.default2), we get std. error
|
|
|
|
Coefficients:
|
|
Estimate Std. Error
|
|
(Intercept) -1.198e+01 6.279e-01
|
|
income 2.885e-05 7.060e-06
|
|
balance 5.822e-03 3.232e-04
|
|
|
|
──────────────────────────────────────────────────────────────────────────
|
|
(b) Write a function, boot.fn() , that takes as input the\
|
|
Default data set as well as an index of the observations, and
|
|
that outputs the coefficient estimates for income and balance in
|
|
the multiple logistic regression model.
|
|
|
|
boot.fn = function(Default,index){
|
|
model = glm(default ~ income + balance,Default,family="binomial",subset=index)
|
|
model$coefficients[2:3]
|
|
}
|
|
|
|
──────────────────────────────────────────────────────────────────────────
|
|
(c) Use the boot() function together with your boot.fn()
|
|
function to estimate the standard errors of the logistic
|
|
regression coefficients for income and balance .
|
|
|
|
> set.seed(56)
|
|
> boot(Default,boot.fn,1000)
|
|
|
|
Bootstrap Statistics :
|
|
original bias std. error
|
|
t1* 2.080898e-05 1.436086e-07 4.679106e-06
|
|
t2* 5.647103e-03 1.876364e-05 2.320547e-04
|
|
|
|
***
|
|
The std. error in t1 above is the std. error for the income
|
|
coefficient, and under t2, the std. error for the balance
|
|
coefficient.
|
|
|
|
──────────────────────────────────────────────────────────────────────────
|
|
(d) Comment on the estimated standard errors obtained using the
|
|
glm() function and using your bootstrap function.
|
|
|
|
|
|
The bootstrap method uses random sampling to approximate
|
|
sampling a real response. This gives a stronger σ²
|
|
(variance) estimate than the formulaic approach used to
|
|
calculate the std err for a linear regression model. The std
|
|
err estimates from the bootstrap method are therefore more
|
|
reliable than those computed earlier in this exercise.
|
|
|
|
The std err from bootstrap for the income coefficient is
|
|
about 66% of that computed from the linear regression model
|
|
directly. This is 72% for the coefficient of balance. These
|
|
smaller values should be considered reliable, that is, we
|
|
can believe the smaller errors compared to the standard
|
|
meothods, because of the reason I explained above. |