轮廓系数(Average silhouette) | 最佳聚类数的判定

news2024/11/14 20:18:41

1.最佳分类个数

# 辅助确定最佳聚类数  4.7*2.6
factoextra::fviz_nbclust( t(DPAU_2), kmeans, method = "silhouette")

在这里插入图片描述
在2有下降拐点，但是样本较多时分成2类一般意义不大。
在7时也有下降拐点。

2.查看每个分类的轮廓系数

(1) pam k=5

library(cluster)
set.seed(101)
pamclu=cluster::pam(t(DPAU_2), k=5)

{
pdf( paste0(outputRoot, keyword, "_01_2.K_means.silhouette.pdf"), width=6, height=5)
df1=silhouette(pamclu)
plot(silhouette(pamclu), 
     col = (1+ head(df1, n=nrow(df1)) |> as.data.frame() |> pull("cluster")),
     main=NULL)
dev.off()
}

在这里插入图片描述
Fig1. Silhouette plot displaying the composition (n = number of samples) and stability (average width) of clustering.

(2) pam k=6

library(cluster)
set.seed(101)
pamclu=cluster::pam(t(DPAU_2), k=6)

{
pdf( paste0(outputRoot, keyword, "_01_2.K_means.6.silhouette.pdf"), width=6, height=5)
df1=silhouette(pamclu)
plot(silhouette(pamclu), 
     col = (1+ head(df1, n=nrow(df1)) |> as.data.frame() |> pull("cluster")),
     main=NULL)
dev.off()
}

在这里插入图片描述

(3) pam k=7

library(cluster)
set.seed(101)
pamclu=cluster::pam(t(DPAU_2), k=7)

{
pdf( paste0(outputRoot, keyword, "_01_2.K_means.7.silhouette.pdf"), width=6, height=5)
df1=silhouette(pamclu)
df1=head(df1, n=nrow(df1)) |> as.data.frame()
plot(silhouette(pamclu), 
     col = df1$cluster +1,
     #xlim=c(min(df1$sil_width)-0.2, max(df1$sil_width))+0.2,
     main=NULL)
dev.off()
}

在这里插入图片描述

(4) kmeans k=5

dat=DPAU_2
kclu=kmeans(t(dat), centers=5)

#kclu$clustering=kclu$cluster #add this list element: clustering
distance=dist( t(dat) )  #10min
kclu.sil=sortSilhouette( silhouette(kclu$cluster, dist = distance ) )
#rownames(kclu.sil)=colnames(dat)
head(kclu.sil)
#     cluster neighbor sil_width
#1226       1        2 0.1124117
#991        1        2 0.1113240
pdf( paste0(outputRoot, keyword, "_01_2.K_means.5.silhouette.pdf"), width=6, height=5)
df1=kclu.sil
df1=head(df1, n=nrow(df1)) |> as.data.frame()
plot(kclu.sil, 
     col = rev(df1$cluster+1),
     do.col.sort=F,
     main=NULL)

factoextra::fviz_silhouette(kclu.sil)
dev.off()

在这里插入图片描述

(5) hclust, k=4

dat=DPAU_2
distance=dist( t(dat) ) 
out.hclust=hclust(distance, method = "ward.D2")

# visual
pdf( paste0(outputRoot, keyword, "_01_2.hclust.4.silhouette.pdf"), width=6, height=5)
plot(out.hclust,
     #hang = -1,
     #hang=0.1,
     hang=0,
     ann=F, axes=F, 
     labels = F, #no labels
     cex = 0.7,
     col = "grey20")
rect.hclust( out.hclust, k=4, border = c("#FF6B6B", "#4ECDC4", "#556270", "deeppink") )
# sil plot
out.hclust.D2=cutree(out.hclust, k=4)
sil_hclust=sortSilhouette(silhouette(out.hclust.D2, distance))
rownames(sil_hclust) = rownames(as.matrix(distance))[attr(sil_hclust, 'iOrd')]
#
plot(sil_hclust, 
     col=out.hclust.D2[rownames( head(sil_hclust, n=nrow(sil_hclust)) )]+1,
     main=attr(sil_hclust, "call") |> deparse() )
dev.off()

在这里插入图片描述