随时间衰减的行总和（rollapply） [英] Sum over rows (rollapply) with time decay

问题描述

这是我刚才发布的问题的后续问题（请参阅对具有多个更改条件的行进行求和R data.table 以获取更多详细信息）。我想计算3个受试者在过去5年中经历过一次活动的次数。因此，使用 zoo 软件包中的 rollapply 这假设5年前的经验和1年前的经验一样重要（相同的权重），所以现在我想包括一个时间衰减的经验，进入总和。这基本上意味着，5年前的经验并没有以与1年前的经验相同的权重来计算总和。

在我的案例中，我想包含一个年龄相关的衰变（即使对于其他应用程序，更快或更慢的衰减，如平方根或方块也是可能的）。

例如，假设我有以下数据（为了清楚起见，我建立在以前的数据基础上）：

  mydf<  -  data.frame（Year = c（2000，2001，2002，2004，2005，
 2007，2000，2001，2002，2003，
 2003 ，2004,2005,2006,2006,2007），
 Name = c（Tom，Tom，Tom，Fred，Gill，
Fred，Gill ，Gill，Gill，Gill，Gill））
 
＃为体验创建一个指标
 mydf $ Ind<  -  1 
 
＃加载需要的包
库（data.table）
库（动物园）
 
＃set data.table 
 setDT（mydf）
 setkey（mydf，Name，Year）
 
＃执行笛卡尔联接来计算经验。 I2是新的经验指示符
m < -  mydf [CJ（unique（Name），seq（min（Year）-5，max（Year））），allow.cartesian = TRUE] [，
列表（Ind = unique（Ind），I2 = sum（Ind，na.rm = TRUE）），
 keyby = list（Name，Year）] 
 
这是方法I已经走了这么远。注意，是一个简单的滚动总和I2 
m [，Exp：= rollapply（I2，5，function（x）sum（head（x，-1）），
 align ='right' = 0），by = Name] 
  

现在问题是，如何包括年龄相关的衰变进入这个计算。为了建模，我需要将经验除以经验的年龄，然后再输入总和。

我一直在努力使用这些行的东西

  m [，Exp_age：= rollapply（I2，5，function（x）sum（head（x，-1） /（tail（（Year）） -  head（Year，-1））），
 align ='right'，fill = 0），by = Name] 
  / pre> 
 
 
但它不工作。我认为我的主要问题是，我不能得到经验的年龄，所以我可以除以年龄的总和。结果应类似于 myres   data.frame 中的 Exp_age  $ c> 
  myres<  -  data.frame（Name = c（Fred，Fred，Fred ，Gill，Gill，Gill，Gill，
Tom，Tom， Tom，Tom，Tom），
 Year = c（2003,2004,2004,2006,2007,2000,2001,2005,2007,2007,2005,2006,2007,2000,2001 ，2002,2002,2003），
 Ind = c（1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1），
 Exp = c（0,1,1,3,4,0,1,1,1,2,3,0,1,2,2,4），
 Exp_age = c（0， 1,1,1,33333333,1.916666667,0,1,0.45，
 0.45,2.2,2,0,1,1.5,1.5,2.833333333））
  
任何指针都非常感谢！
解决方案
正确地，您试图用 width = 5 做一个 rollapply ，而不是做一个简单的和，你想做加权和。权重是相对于5年窗口的经验的年龄。我会这样做：首先设置你的 data.table 中的键，使它具有适当的增加顺序 Name 那么你知道你的 x 变量中的最后一个项目是最小的，第一个项目是最旧的（你已经在你的代码中这样做）。我不能完全知道你想要的权重去哪儿（最小的有最大的重量或最老的），但你得到的点：
  setkey（m，Name，Year）
 my_fun = function（x）{w = 1：length（x）; sum（x * w）} 
m [，Exp_age：= rollapply（I2，width = 5，by = 1，fill = NA，FUN = my_fun，by.column = FALSE，align =right） = Name] 
  
 
This is a follow on question to a question I posted earlier (see Sum over rows with multiple changing conditions R data.table for more details). I want to calculate how many times the 3 subjects have experienced an event in the last 5 years. So have been summing over a rolling window using rollapply from the zoo package. This assumes that the experience 5 years ago is as important as the experience 1 year ago (same weighting), so now I want to include a time decay for the experience that enters the sum. This basically means that the experience 5 years ago does not enter into the sum with the same weighting as the experience 1 year ago. 


I my case I want to include an age dependent decay (even though for other applications faster or slower decays such as square root or squares could be possible). 

For example lets assume I have the following data (I build on the previous data for clarity): 
mydf <- data.frame (Year = c(2000, 2001, 2002, 2004, 2005,
                         2007, 2000, 2001, 2002, 2003,
                         2003, 2004, 2005, 2006, 2006, 2007),
                Name = c("Tom", "Tom", "Tom", "Fred", "Gill",
                         "Fred", "Gill", "Gill", "Tom", "Tom",
                         "Fred", "Fred", "Gill", "Fred", "Gill", "Gill"))

# Create an indicator for the experience 
mydf$Ind <- 1

# Load require packages
library(data.table)
library(zoo)

# Set data.table
setDT(mydf)
setkey(mydf, Name,Year)

# Perform cartesian join to calculate experience. I2 is the new experience indicator 
m <- mydf[CJ(unique(Name),seq(min(Year)-5, max(Year))),allow.cartesian=TRUE][,
        list(Ind = unique(Ind), I2 = sum(Ind,na.rm=TRUE)),
        keyby=list(Name,Year)]

# This is the approach I have been taking so far. Note that is a simple rolling sum of I2
m[,Exp := rollapply(I2, 5, function(x) sum(head(x,-1)), 
                align = 'right', fill=0),by=Name]
So question now is, how can I include a age dependent decay into this calculation. To model this I need to divide the experience by the age of the experience before it enters the sum.

I have been trying to get it to work using something along these lines:
 m[,Exp_age := rollapply(I2, 5, function(x) sum(head(x,-1)/(tail((Year))-head(Year,-1))), 
                     align = 'right', fill=0),by=Name]
But it does not work. I think my main problem is that I cannot get the age of the experience right so I can divide by the age in the sum. The result should look like the Exp_age column in the myres data.frame below
myres <- data.frame(Name = c("Fred", "Fred", "Fred", "Fred", "Fred", 
                         "Gill", "Gill", "Gill", "Gill", "Gill", "Gill", 
                         "Tom", "Tom", "Tom", "Tom", "Tom"), 
                Year = c(2003, 2004, 2004, 2006, 2007, 2000, 2001, 2005,
                         2005, 2006, 2007, 2000, 2001, 2002, 2002, 2003), 
                Ind = c(1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1), 
                Exp = c(0, 1, 1, 3, 4, 0, 1, 1, 1, 2, 3, 0, 1, 2, 2, 4), 
                Exp_age = c(0, 1, 1, 1.333333333, 1.916666667, 0, 1, 0.45, 
                            0.45, 2.2, 2, 0, 1, 1.5, 1.5, 2.833333333))
Any pointers would be greatly appreciated!
解决方案
If I understand you correctly, you are trying to do a rollapply with width=5 and rather than do a simple sum, you want to do a weighted sum. The weights are the age of the experience relative to the 5 year window. I would do this: first set the key in your data.table so that it has proper increasing order by Name, then you know that the last item in your x variable is the youngest and the first item is the oldest (you do this in your code already). I can't quite tell which way you want the weights to go (youngest to have greatest weight or oldest) but you get the point:
setkey(m, Name, Year)
my_fun = function(x) { w = 1:length(x); sum(x*w)}
m[,Exp_age:=rollapply(I2, width=5, by=1, fill=NA, FUN=my_fun, by.column=FALSE, align="right") ,by=Name]


                        
这篇关于随时间衰减的行总和（rollapply）的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！