apfp: Model-based Clustering with APFP

Description Usage Arguments Details Value References See Also Examples

View source: R/apfp.R

Description

The adaptive pairwise fusion penalty (APFP) was proposed by Guo (2010). Under the framework of the model-based clustering, APFP aims to identify the pairwise informative variables for clustering high-dimenisonal data.

Usage

1
2
3
4
5
6
apfp(tuning, K = NULL, lambda = NULL, y, N = 100, kms.iter = 100, kms.nstart = 100,
      adapt.kms = FALSE, eps.diff = 1e-5, eps.em = 1e-5,
      iter.LQA = 20, eps.LQA = 1e-5, model.crit = 'gic')
apfp(tuning = NULL, K, lambda, y, N = 100, kms.iter = 100, kms.nstart = 100,
      adapt.kms = FALSE, eps.diff = 1e-5, eps.em = 1e-5,
      iter.LQA = 20, eps.LQA = 1e-5, model.crit = 'gic')

Arguments

tuning

A 2-dimensional vector or a matrix with 2 columns, the first column is the number of clusters K and the second column is the tuning parameter λ in the penalty term. If this is missing, then K and lambda must be provided.

K

The number of clusters K.

lambda

The tuning parameter λ in the penalty term.

y

A p-dimensional data matrix. Each row is an observation.

N

The maximum number of iterations in the EM algorithm. The default value is 100.

kms.iter

The maximum number of iterations in kmeans algorithm for generating the starting value for the EM algorithm.

kms.nstart

The number of starting values in K-means.

adapt.kms

A indicator of using the cluster means estimated by K-means to calculate the adaptive parameters in APFP. The default value is FALSE.

eps.diff

The lower bound of pairwise difference of two mean values. Any value lower than it is treated as 0.

eps.em

The lower bound for the stopping criterion in the EM algorithm.

iter.LQA

The number of iterations in the estimation of cluster means by using the local quadratic approximation (LQA).

eps.LQA

The lower bound for the stopping criterion in the estimation of cluster means.

model.crit

The criterion used to select the number of clusters K. It is either ‘bic’ for Bayesian Information Criterion or ‘gic’ for Generalized Information Criterion.

Details

The j-th variable is defined as pairwise informative for a pair of clusters C_k and C_{k'} if μ_{kj} \neq μ_{k'j}. Also, a variable is globally informative if it is pairwise informative for at least one pair of clusters. Here we assume that each cluster has the same diagonal variance in the model-based clustering. APFP is in the following form,

∑_{j=1}^d ∑_{k<k'}τ_{kk'j}|μ_{kj} - μ_{k'j}|,

where d is the number of variables in the data, τ_{kk'j} = |\tilde{μ}_{kj} - \tilde{μ}_{k'j}|^{-1} is the adaptive parameters. Here we provide two choices for \tilde{μ_{kj}}. If adapt.kms == TRUE, \tilde{μ}_{kj} is the estimates from the K-mean algorithm; otherwise, \tilde{μ}_{kj} is the estimates from the model-based clustering without penalty.

The estimation uses the EM algorithm. Since the EM algorithm depends on the starting values. We use the estimates from K-means with multiple starting points as the starting values. For estimating the cluster means, APFP uses the local quadratic approximation.

Value

This function returns the esimated parameters and some statistics of the optimal model within the given K and λ, which is selected by BIC when model.crit = 'bic' or GIC when model.crit = 'gic'.

mu.hat.best

The estimated cluster means in the optimal model

sigma.hat.best

The estimated covariance in the optimal model

p.hat.best

The estimated cluster proportions in the optimal model

s.hat.best

The clustering assignments using the optimal model

lambda.best

The value of λ that provide the optimal model

K.best

The value of K that provide the optimal model

llh.best

The log-likelihood of the optimal model

gic.best

The GIC of the optimal model

bic.best

The BIC of the optimal model

ct.mu.best

The degrees of freedom in the cluster means of the optimal model

References

Guo, J., Levina, E., Michailidis, G., and Zhu, J. (2010) Pairwise variable selection for high-dimensional model-based clustering. Biometrics 66(3), 793–804.

See Also

nopenalty apL1 parse

Examples

1
2
3
y <- rbind(matrix(rnorm(100,0,1),ncol=2), matrix(rnorm(100,4,1), ncol=2))
output <- apfp(K = c(1:2), lambda = c(0,1), y=y)
output$mu.hat.best

PARSE documentation built on May 30, 2017, 1:16 a.m.