虽然我喜欢肯威廉姆斯简单的功能，我想检索多种模式，如果他们存在。考虑到这一点，我使用下面的函数，它返回多个模式或单个模式的列表。

rmode <- function(x) {
  x <- sort(x)  
  u <- unique(x)
  y <- lapply(u, function(y) length(x[x==y]))
  u[which( unlist(y) == max(unlist(y)) )]
} 

我将使用density()函数来确定一个(可能是连续的)分布的平滑最大值:

function(x) density(x, 2)$x[density(x, 2)$y == max(density(x, 2)$y)]

其中x是数据集合。注意调节平滑的密度函数的调节参数。

假设你的观测值是来自实数的类，当你的观测值是2,2,3,3时，你期望模态为2.5，然后你可以用mode = l1 + I * (f1-f0) / (2f1 -f0 - f2)来估计模态，其中l1..最频繁类的下限，f1..最频繁类的频率，f0..在最频繁类之前的类的频率，f2..在最频繁类之后的类的频率，i..分类间隔，如在1,2,3中给出:

#Small Example
x <- c(2,2,3,3) #Observations
i <- 1          #Class interval

z <- hist(x, breaks = seq(min(x)-1.5*i, max(x)+1.5*i, i), plot=F) #Calculate frequency of classes
mf <- which.max(z$counts)   #index of most frequent class
zc <- z$counts
z$breaks[mf] + i * (zc[mf] - zc[mf-1]) / (2*zc[mf] - zc[mf-1] - zc[mf+1])  #gives you the mode of 2.5


#Larger Example
set.seed(0)
i <- 5          #Class interval
x <- round(rnorm(100,mean=100,sd=10)/i)*i #Observations

z <- hist(x, breaks = seq(min(x)-1.5*i, max(x)+1.5*i, i), plot=F)
mf <- which.max(z$counts)
zc <- z$counts
z$breaks[mf] + i * (zc[mf] - zc[mf-1]) / (2*zc[mf] - zc[mf-1] - zc[mf+1])  #gives you the mode of 99.5

如果你想要最频繁的级别，并且你有多个最频繁的级别，你可以得到所有的级别，例如:

x <- c(2,2,3,5,5)
names(which(max(table(x))==table(x)))
#"2" "5"

我浏览了所有这些选项，开始想知道它们的相对特性和性能，所以我做了一些测试。如果其他人也好奇，我在这里分享我的结果。

我不想为这里发布的所有函数而烦恼，我选择了一个基于一些标准的示例:函数应该对字符、因子、逻辑和数字向量都有效，它应该适当地处理na和其他有问题的值，输出应该是“合理的”，即没有数字作为字符或其他类似的愚蠢行为。

我还添加了一个我自己的函数，它是基于与chrispy相同的想法，除了适应更一般的用途:

library(magrittr)

Aksel <- function(x, freq=FALSE) {
    z <- 2
    if (freq) z <- 1:2
    run <- x %>% as.vector %>% sort %>% rle %>% unclass %>% data.frame
    colnames(run) <- c("freq", "value")
    run[which(run$freq==max(run$freq)), z] %>% as.vector   
}

set.seed(2)

F <- sample(c("yes", "no", "maybe", NA), 10, replace=TRUE) %>% factor
Aksel(F)

# [1] maybe yes  

C <- sample(c("Steve", "Jane", "Jonas", "Petra"), 20, replace=TRUE)
Aksel(C, freq=TRUE)

# freq value
#    7 Steve

最后，我通过微基准测试在两组测试数据上运行了五个函数。函数名指的是它们各自的作者:

Chris的函数被设置为method="modes"和na。rm=TRUE默认值，以使其更具可比性，但除此之外，这里使用的函数是由它们的作者提供的。

In matter of speed alone Kens version wins handily, but it is also the only one of these that will only report one mode, no matter how many there really are. As is often the case, there's a trade-off between speed and versatility. In method="mode", Chris' version will return a value iff there is one mode, else NA. I think that's a nice touch. I also think it's interesting how some of the functions are affected by an increased number of unique values, while others aren't nearly as much. I haven't studied the code in detail to figure out why that is, apart from eliminating logical/numeric as a the cause.

这里有另一个解决方案:

freq <- tapply(mySamples,mySamples,length)
#or freq <- table(mySamples)
as.numeric(names(freq)[which.max(freq)])

基于@Chris的函数来计算模态或相关指标，但是使用Ken Williams的方法来计算频率。这个方法修复了根本没有模式(所有元素频率相等)的情况，并提供了一些更易读的方法名。

Mode <- function(x, method = "one", na.rm = FALSE) {
  x <- unlist(x)
  if (na.rm) {
    x <- x[!is.na(x)]
  }

  # Get unique values
  ux <- unique(x)
  n <- length(ux)

  # Get frequencies of all unique values
  frequencies <- tabulate(match(x, ux))
  modes <- frequencies == max(frequencies)

  # Determine number of modes
  nmodes <- sum(modes)
  nmodes <- ifelse(nmodes==n, 0L, nmodes)

  if (method %in% c("one", "mode", "") | is.na(method)) {
    # Return NA if not exactly one mode, else return the mode
    if (nmodes != 1) {
      return(NA)
    } else {
      return(ux[which(modes)])
    }
  } else if (method %in% c("n", "nmodes")) {
    # Return the number of modes
    return(nmodes)
  } else if (method %in% c("all", "modes")) {
    # Return NA if no modes exist, else return all modes
    if (nmodes > 0) {
      return(ux[which(modes)])
    } else {
      return(NA)
    }
  }
  warning("Warning: method not recognised.  Valid methods are 'one'/'mode' [default], 'n'/'nmodes' and 'all'/'modes'")
}

由于它使用Ken的方法来计算频率，性能也得到了优化，使用AkselA的帖子，我对之前的一些答案进行了基准测试，以显示我的函数在性能上是如何接近Ken的，各种输出选项的条件只导致很小的开销: