问题
I know that in ggplot2 one can add the convex hull to a scatterplot by group as in
library(ggplot2)
library(plyr)
data(iris)
df<-iris
find_hull <- function(df) df[chull(df$Sepal.Length, df$Sepal.Width), ]
hulls <- ddply(df, "Species", find_hull)
plot <- ggplot(data = df, aes(x = Sepal.Length, y = Sepal.Width, colour=Species, fill = Species)) +
geom_point() +
geom_polygon(data = hulls, alpha = 0.5) +
labs(x = "Sepal.Length", y = "Sepal.Width")
plot
I was wondering though how one could calculate and add alpha bags instead, i.e. the largest convex hull that contains at least a proportion 1-alpha of all the points? Either in 2d (to display with ggplot2) or 3d (to display with rgl).
EDIT: My initial idea was be to keep on "peeling" the convex hull for as along as the criterion of containing at least a given % of points would be satisfied, although in the paper here it seems they use a different algorithm (isodepth, which seems to be implemented in R package depth, in function isodepth and aplpack::plothulls seems also close to what I want (although it produces a full plot as opposed to just the contour), so I think with these I may be sorted. Though these function only works in 2D, and I would also be interested in a 3D extension (to be plotted in rgl). If anyone has any pointers let me know!
EDIT2: with function depth::isodepth I found a 2d solution (see post below), although I am still looking for a 3D solution as well - if anyone would happen to know how to do that, please let me know!
回答1:
We can modify the aplpack::plothulls function to accept a parameter for the proportion of points to enclose (in aplpack it's set to 50%). Then we can use this modified function to make a custom a geom for ggplot.
Here's the custom geom:
library(ggplot2)
StatBag <- ggproto("Statbag", Stat,
compute_group = function(data, scales, prop = 0.5) {
#################################
#################################
# originally from aplpack package, plotting functions removed
plothulls_ <- function(x, y, fraction, n.hull = 1,
col.hull, lty.hull, lwd.hull, density=0, ...){
# function for data peeling:
# x,y : data
# fraction.in.inner.hull : max percentage of points within the hull to be drawn
# n.hull : number of hulls to be plotted (if there is no fractiion argument)
# col.hull, lty.hull, lwd.hull : style of hull line
# plotting bits have been removed, BM 160321
# pw 130524
if(ncol(x) == 2){ y <- x[,2]; x <- x[,1] }
n <- length(x)
if(!missing(fraction)) { # find special hull
n.hull <- 1
if(missing(col.hull)) col.hull <- 1
if(missing(lty.hull)) lty.hull <- 1
if(missing(lwd.hull)) lwd.hull <- 1
x.old <- x; y.old <- y
idx <- chull(x,y); x.hull <- x[idx]; y.hull <- y[idx]
for( i in 1:(length(x)/3)){
x <- x[-idx]; y <- y[-idx]
if( (length(x)/n) < fraction ){
return(cbind(x.hull,y.hull))
}
idx <- chull(x,y); x.hull <- x[idx]; y.hull <- y[idx];
}
}
if(missing(col.hull)) col.hull <- 1:n.hull
if(length(col.hull)) col.hull <- rep(col.hull,n.hull)
if(missing(lty.hull)) lty.hull <- 1:n.hull
if(length(lty.hull)) lty.hull <- rep(lty.hull,n.hull)
if(missing(lwd.hull)) lwd.hull <- 1
if(length(lwd.hull)) lwd.hull <- rep(lwd.hull,n.hull)
result <- NULL
for( i in 1:n.hull){
idx <- chull(x,y); x.hull <- x[idx]; y.hull <- y[idx]
result <- c(result, list( cbind(x.hull,y.hull) ))
x <- x[-idx]; y <- y[-idx]
if(0 == length(x)) return(result)
}
result
} # end of definition of plothulls
#################################
# prepare data to go into function below
the_matrix <- matrix(data = c(data$x, data$y), ncol = 2)
# get data out of function as df with names
setNames(data.frame(plothulls_(the_matrix, fraction = prop)), nm = c("x", "y"))
# how can we get the hull and loop vertices passed on also?
},
required_aes = c("x", "y")
)
#' @inheritParams ggplot2::stat_identity
#' @param prop Proportion of all the points to be included in the bag (default is 0.5)
stat_bag <- function(mapping = NULL, data = NULL, geom = "polygon",
position = "identity", na.rm = FALSE, show.legend = NA,
inherit.aes = TRUE, prop = 0.5, alpha = 0.3, ...) {
layer(
stat = StatBag, data = data, mapping = mapping, geom = geom,
position = position, show.legend = show.legend, inherit.aes = inherit.aes,
params = list(na.rm = na.rm, prop = prop, alpha = alpha, ...)
)
}
geom_bag <- function(mapping = NULL, data = NULL,
stat = "identity", position = "identity",
prop = 0.5,
alpha = 0.3,
...,
na.rm = FALSE,
show.legend = NA,
inherit.aes = TRUE) {
layer(
data = data,
mapping = mapping,
stat = StatBag,
geom = GeomBag,
position = position,
show.legend = show.legend,
inherit.aes = inherit.aes,
params = list(
na.rm = na.rm,
alpha = alpha,
prop = prop,
...
)
)
}
#' @rdname ggplot2-ggproto
#' @format NULL
#' @usage NULL
#' @export
GeomBag <- ggproto("GeomBag", Geom,
draw_group = function(data, panel_scales, coord) {
n <- nrow(data)
if (n == 1) return(zeroGrob())
munched <- coord_munch(coord, data, panel_scales)
# Sort by group to make sure that colors, fill, etc. come in same order
munched <- munched[order(munched$group), ]
# For gpar(), there is one entry per polygon (not one entry per point).
# We'll pull the first value from each group, and assume all these values
# are the same within each group.
first_idx <- !duplicated(munched$group)
first_rows <- munched[first_idx, ]
ggplot2:::ggname("geom_bag",
grid:::polygonGrob(munched$x, munched$y, default.units = "native",
id = munched$group,
gp = grid::gpar(
col = first_rows$colour,
fill = alpha(first_rows$fill, first_rows$alpha),
lwd = first_rows$size * .pt,
lty = first_rows$linetype
)
)
)
},
default_aes = aes(colour = "NA", fill = "grey20", size = 0.5, linetype = 1,
alpha = NA, prop = 0.5),
handle_na = function(data, params) {
data
},
required_aes = c("x", "y"),
draw_key = draw_key_polygon
)
And here's an example of how it can be used:
ggplot(iris, aes(Sepal.Length, Petal.Length, colour = Species, fill = Species)) +
geom_point() +
stat_bag(prop = 0.95) + # enclose 95% of points
stat_bag(prop = 0.5, alpha = 0.5) + # enclose 50% of points
stat_bag(prop = 0.05, alpha = 0.9) # enclose 5% of points
回答2:
Ha with the help of function depth::isodepth I came up with the following solution - here I find the alpha bag that contains a proportion of at least 1-alpha of all points :
library(mgcv)
library(depth)
library(plyr)
library(ggplot2)
data(iris)
df=iris[,c(1,2,5)]
alph=0.05
find_bag = function(x,alpha=alph) {
n=nrow(x)
target=1-alpha
propinside=1
d=1
while (propinside>target) {
p=isodepth(x[,1:2],dpth=d,output=T, mustdith=T)[[1]]
ninside=sum(in.out(p,as.matrix(x[,1:2],ncol=2))*1)
nonedge=sum(sapply(1:nrow(p),function (row)
nrow(merge(round(setNames(as.data.frame(p[row,,drop=F]),names(x)[1:2]),5),as.data.frame(x[,1:2])))>0)*1)-3
propinside=(ninside+nonedge)/n
d=d+1
}
p=isodepth(x[,1:2],dpth=d-1,output=T, mustdith=T)[[1]]
p }
bags <- ddply(df, "Species", find_bag,alpha=alph)
names(bags) <- c("Species",names(df)[1:2])
plot <- ggplot(data = df, aes(x = Sepal.Length, y = Sepal.Width, colour=Species, fill = Species)) +
geom_point() +
geom_polygon(data = bags, alpha = 0.5) +
labs(x = "Sepal.Length", y = "Sepal.Width")
plot
EDIT2: Using my original idea of convex hull peeling I also came up with the following solution which now works in 2d & 3d; the result is not quite the same is with the isodepth algorithm, but it's pretty close :
# in 2d
library(plyr)
library(ggplot2)
data(iris)
df=iris[,c(1,2,5)]
alph=0.05
find_bag = function(x,alpha=alph) {
n=nrow(x)
propinside=1
target=1-alpha
x2=x
while (propinside>target) {
propinside=nrow(x2)/n
hull=chull(x2)
x2old=x2
x2=x2[-hull,]
}
x2old[chull(x2old),] }
bags <- ddply(df, "Species", find_bag, alpha=alph)
plot <- ggplot(data = df, aes(x = Sepal.Length, y = Sepal.Width, colour=Species, fill = Species)) +
geom_point() +
geom_polygon(data = bags, alpha = 0.5) +
labs(x = "Sepal.Length", y = "Sepal.Width")
plot
# in 3d
library(plyr)
library(ggplot2)
data(iris)
df=iris[,c(1,2,3,5)]
levels=unique(df[,"Species"])
nlevels=length(levels)
zoom=0.8
cex=1
aspectr=c(1,1,0.7)
pointsalpha=1
userMatrix=matrix(c(0.80,-0.60,0.022,0,0.23,0.34,0.91,0,-0.55,-0.72,0.41,0,0,0,0,1),ncol=4,byrow=T)
windowRect=c(0,29,1920,1032)
cols=c("red","forestgreen","blue")
alph=0.05
plotbag = function(x,alpha=alph,grp=1,cols=c("red","forestgreen","blue"),transp=0.2) {
propinside=1
target=1-alpha
x2=x
levels=unique(x2[,ncol(x2)])
x2=x2[x2[,ncol(x2)]==levels[[grp]],]
n=nrow(x2)
while (propinside>target) {
propinside=nrow(x2)/n
hull=unique(as.vector(convhulln(as.matrix(x2[,1:3]), options = "Tv")))
x2old=x2
x2=x2[-hull,]
}
ids=t(convhulln(as.matrix(x2old[,1:3]), options = "Tv"))
rgl.triangles(x2old[ids,1],x2old[ids,2],x2old[ids,3],col=cols[[grp]],alpha=transp,shininess=50)
}
open3d(zoom=zoom,userMatrix=userMatrix,windowRect=windowRect,antialias=8)
for (i in 1:nlevels) {
plot3d(x=df[df[,ncol(df)]==levels[[i]],][,1],
y=df[df[,ncol(df)]==levels[[i]],][,2],
z=df[df[,ncol(df)]==levels[[i]],][,3],
type="s",
col=cols[[i]],
size=cex,
lit=TRUE,
alpha=pointsalpha,point_antialias=TRUE,
line_antialias=TRUE,shininess=50, add=TRUE)
plotbag(df,alpha=alph, grp=i, cols=c("red","forestgreen","blue"), transp=0.3) }
axes3d(color="black",drawfront=T,box=T,alpha=1)
title3d(color="black",xlab=names(df)[[1]],ylab=names(df)[[2]],zlab=names(df)[[3]],alpha=1)
aspect3d(aspectr)
来源：https://stackoverflow.com/questions/31893559/r-adding-alpha-bags-to-a-2d-or-3d-scatterplot