This package includes the following parts. First, Three methods (extended gap statistic, extedned prediction strength and S4) for estimating K and lambda simultaneously in sparse K-means are included in KL.Gap, KL.PS and KL.S4 respectively. Second, 11 methods for estimating K in K-means are included in K.Clust function. Third, 9 high-dimensional datasets are also included with name starting with ds. Several simulation functions are included in Sim1, Sim2 and Sim3. The extensive simulation settings and real applications in this paper can also serve as a standard benchmark for evaluation purpose when new methods for simultaneous estimation of K and lambda are developed. Finally, an efficient algorithm for generate lambda grid for each K is included in the region.lambda.
S4_package_example is a R script for demo purpose. The zip file data after preprocessing has all the preprocessed data, whcih are also imported in the R package S4. The Rscripts folder contains all the code to reproduce the result in the following paper:
- Li, Yujia, et al. "Simultaneous Estimation of Number of Clusters and Feature Sparsity Parameter in High-dimensional clustering analysis". Biometrics, accepted (2021). A previous arxiv version: arXiv:1909.01930
S4 package files are in S4/ folder, You can install by copying and paste the following code into R
devtools::install_github("YujiaLi1994/S4/S4")
Alternatively, download the tar.gz zipped file, unzip it and install ogClust package using devtools::install(), which can make sure to install all the dependences. Make sure R packages mvtnorm, caret, parallel, irr, plyr,cluster, fpc, MCMCpack, MASS, sparcl have all been properly imported.
#simulate the data of simulation I by Sim1 function.
data<-Sim1(settings = "A1")#A1-A7 are well separated data and B1-B7 are not-well separated data
#using S4
#trim.S4 is the percentage of trimmed scattered points.
#Cutoff is the threshold to detect the K=1
#n.resample is number of subsampling
res.S4<-K.Clust(data$x,method="S4",Kmin=2,Kmax=10,trim.S4=0.05,cutoff=0.8,n.resample=50)
#users can also implement other methods, see the document and reference for details.
#Take prediction strength for an example
res.PS<-K.Clust(data$x,method="PS",Kmin=2,Kmax=10,trim.S4=0.05,cutoff=0.8,n.resample=50)- Simulate a dataset using Sim2 function.
data<-Sim2(h=200,q = 50,u=0.8)#For demo purpose, we use 200 features in total for fast result.
k_vector<-2:7#K is selected from 2 to 7- Use Efficient Algorithm for Choosing Grids of lambda for each K
wbounds_list<-list(1)
for(l in 1:length(k_vector)){#for each K, using the algorithm to get 20 lambda.
wbounds_list[[l]] = region.lambda(lam1=1.5,iteration=20,data,k_vector[l])
}- Delete lambda which will select all the feature, since our S4 method calculate specificity, the lambda selecting all the genes must be removed. Including very large lambda (that select all feature) has the possibility to cause an error.
for(l in 1:length(k_vector)){
temp<-KMeansSparseCluster(data,K=k_vector[l],wbounds=wbounds_list[[l]],nstart=100)
num<-rep(0,length(temp))
for(i in 1:length(num)){#get the corresponding number of features for each K and each lambda
num[i]<-sum(temp[[i]]$ws>0)
}
if(sum(num==ncol(data))>0){#For each K, if a certain lambda selects all features, delete it.
#For a large simulation study, two closest lambda next to it can also be removed to be conservative.
wbounds_list[[l]]<-wbounds_list[[l]][1:(min(which(num==ncol(data)))-3)]
}
}- implementing S4
#run S4 method
res.S4<-KL.S4(x=data,lambda_list = wbounds_list,k_vector = k_vector,trim =0.05,n.resample = 50,num.cores = 1)- Implement extended Gap and extended Prediction strength
#run extended Gap statistic method
res.Gap<-KL.Gap(x=data,lambda_list = wbounds_list,k_vector = k_vector,n.perm = 50,num.cores = 1)
#run extended Prediction strength method
res.PS<-KL.PS(x=data,lambda_list = wbounds_list,k_vector = k_vector,cv = 2,M=20,num.cores = 1,cutoff = 0.8)All the datasets are imported and accessible after library(S4).
#Data sets in package ‘S4’:
# ds.GSE13159 Microarray data of Leukemia after preprocessing(GSE13159)
# ds.GSE17855 Microarray data of leukemia after preprocessing(GSE17855)
# ds.GSE47474 RNA Sequencing data of rat brain after
# ds.GSE6891 Microarray data of leukemia after preprocessing(GSE6891)
# ds.ISOLET letter-name dataset after preprocessing
# ds.Leaf Plant species leaves dataset after preprocessing
# ds.Pancancer RNA Sequencing data after preprocessing(Pancancer)
# ds.SNP SNP data after preprocessing
# ds.TissueType Microarray data after preprocessing(Mammalian tissue
# types dataset:)