customizedGlmnet {customizedTraining}

fit glmnet using customized training

Description

Fit a regularized lasso model using customized training

Usage

customizedGlmnet(xTrain, yTrain, xTest, groupid = NULL, G = NULL,
    family = c("gaussian", "binomial", "multinomial"), dendrogram = NULL,
    dendrogramTestIndices = NULL)

Arguments

xTrain

an n-by-p matrix of training covariates

yTrain

a length-n vector of training responses. Numeric for family = "gaussian". Factor or character for family = "binomial" or family = "multinomial"

xTest

an m-by-p matrix of test covariates

groupid

an optional length-m vector of group memberships for the test set. If specified, customized training subsets are identified using the union of nearest neighbor sets for each test group. Either groupid or G must be specified

G

a positive integer indicating the number of clusters for the joint clustering of the test and training data. Ignored if groupid is specified. Either groupid or G must be specified

family

response type

dendrogram

optional output from hclust on the joint covariate data. Used by cv.customizedGlmnet so that clustering is not computed redundantly

dendrogramTestIndices

optional set of indices (corresponding to dendrogram) held out in cross-validation. Used by cv.customizedGlmnet

Details

Identify customized training subsets of the training data through one of two methods: (1) If groupid is specified, grouping the test data, then for each test group find the 10 nearest neighbors of each observation in the group and use the union of these nearest neighbor sets as the customized training set or (2) If G is specified, jointly cluster the test and training data using hierarchical clustering with complete linkage. Within each cluster, the training data are used as the customized training subset for the test data. Once the customized training subsets have been identified, use glmnet to fit an l1-regularized regression model to each.

Value

an object with class customizedGlmnet

call

the call that produced this object

CTset

a list containing the customized training subsets for each test group

fit

a list containing the glmnet fit for each test group

groupid

a length-m vector containing the group memberships of the test data

x

a list containing train (which is the input xTrain) and test (which is the input xTest). Specified in function call

y

training response vector (specified in function call)

family

response type (specified in function call)

standard

the fit of glmnet to the entire training set using standard training

Author(s)

Scott Powers, Trevor Hastie, Robert Tibshirani

References

Scott Powers, Trevor Hastie and Robert Tibshirani (2015) "Customized training with an application to mass specrometric imaging of gastric cancer data." Annals of Applied Statistics 9, 4:1709-1725.

See Also

print.customizedGlmnet, predict.customizedGlmnet, plot.customizedGlmnet, cv.customizedGlmnet

Examples

require(glmnet)

# Simulate synthetic data

n = m = 150
p = 50
q = 5
K = 3
sigmaC = 10
sigmaX = sigmaY = 1
set.seed(5914)

beta = matrix(0, nrow = p, ncol = K)
for (k in 1:K) beta[sample(1:p, q), k] = 1
c = matrix(rnorm(K*p, 0, sigmaC), K, p)
eta = rnorm(K)
pi = (exp(eta)+1)/sum(exp(eta)+1)
z = t(rmultinom(m + n, 1, pi))
x = crossprod(t(z), c) + matrix(rnorm((m + n)*p, 0, sigmaX), m + n, p)
y = rowSums(z*(crossprod(t(x), beta))) + rnorm(m + n, 0, sigmaY)

x.train = x[1:n, ]
y.train = y[1:n]
x.test = x[n + 1:m, ]
y.test = y[n + 1:m]


# Example 1: Use clustering to fit the customized training model to training
# and test data with no predefined test-set blocks

fit1 = customizedGlmnet(x.train, y.train, x.test, G = 3,
    family = "gaussian")

# Print the customized training model fit:
fit1

# Extract nonzero regression coefficients for each group:
nonzero(fit1, lambda = 10)

# Compute test error using the predict function:
mean((y.test - predict(fit1, lambda = 10))^2)

# Plot nonzero coefficients by group:
plot(fit1, lambda = 10)


# Example 2: If the test set has predefined blocks, use these blocks to define
# the customized training sets, instead of using clustering.
group.id = apply(z == 1, 1, which)[n + 1:m]

fit2 = customizedGlmnet(x.train, y.train, x.test, group.id)

# Print the customized training model fit:
fit2

# Extract nonzero regression coefficients for each group:
nonzero(fit2, lambda = 10)

# Compute test error using the predict function:
mean((y.test - predict(fit2, lambda = 10))^2)

# Plot nonzero coefficients by group:
plot(fit2, lambda = 10)