变异期间递归滞后（）？

Question

替代标题可能是“使用变异中的滞后来引用先前的行变异”

\n

我想包含为前一行生成的值作为变异计算的输入。一些数据：

\n

mydiamonds <- diamonds %>%\n  mutate(Ideal = ifelse(cut == 'Ideal', 1, 0)) %>% \n  group_by(Ideal) %>% \n  mutate(rn = row_number()) %>% \n  arrange(Ideal, rn) %>% \n  mutate(CumPrice = cumsum(price)) %>% \n  mutate(InitialPrice = min(price)) %>% \n  select(Ideal, rn, CumPrice, InitialPrice)\n

\n

看起来像这样：

\n

mydiamonds %>% head\n# A tibble: 6 x 4\n# Groups:   Ideal [1]\n  Ideal    rn CumPrice InitialPrice\n  <dbl> <int>    <int>        <int>\n1     0     1      326          326\n2     0     2      653          326\n3     0     3      987          326\n4     0     4     1322          326\n5     0     5     1658          326\n6     0     6     1994          326\n

\n

一个模型：

\n

mod.diamonds = glm(CumPrice ~ log(lag(CumPrice)) +log(rn) + Ideal , family = "poisson", data = mydiamonds)\n

\n

测试模型：

\n

# new data, pretend we don't know CumPrice but want to use predictions to predict subsequent predictions\nmydiamonds.testdata <- mydiamonds %>% select(-CumPrice)\n# manual prediction based on lag(prediction), for the first row in each group use InitialPrice\n## add coefficients as fields\ncoeffs <- mod.diamonds$coefficients\nmydiamonds.testdata <- mydiamonds.testdata %>% \n  mutate(CoefIntercept = coeffs['(Intercept)'],\n         CoefLogLagCumPrice = coeffs['log(lag(CumPrice))'],\n         CoefLogRn = coeffs['log(rn)'],\n         CoefIdeal = coeffs['Ideal']\n         )\n

\n

这是我的测试数据的样子：

\n

 mydiamonds.testdata %>% head\n# A tibble: 6 x 7\n# Groups:   Ideal [1]\n  Ideal    rn InitialPrice CoefIntercept CoefLogLagCumPrice CoefLogRn CoefIdeal\n  <dbl> <int>        <int>         <dbl>              <dbl>     <dbl>     <dbl>\n1     0     1          326        0.0931              0.987    0.0154 -0.000715\n2     0     2          326        0.0931              0.987    0.0154 -0.000715\n3     0     3          326        0.0931              0.987    0.0154 -0.000715\n4     0     4          326        0.0931              0.987    0.0154 -0.000715\n5     0     5          326        0.0931              0.987    0.0154 -0.000715\n6     0     6          326        0.0931              0.987    0.0154 -0.000715\n

\n

无法使用 Predict()，因为我需要递归地预测前一天/行的预测输入到当天的位置。相反，尝试使用系数进行手动预测：

\n

# prediction\nmydiamonds.testdata <- mydiamonds.testdata %>% \n  mutate(\n    Prediction = CoefIntercept + \n      \n      # here's the hard bit. If it's the first row in the group, use InitialPrice, else use the value of the previous prediction\n      (CoefLogLagCumPrice * ifelse(rn == 1, InitialPrice, lag(Prediction))) + \n      \n      (CoefLogRn * log(rn)) + \n      (CoefIdeal * Ideal)\n    )\n

\n

\n

错误：mutate()输入有问题Prediction。x 对象\n未找到“预测”\xe2\x84\xb9 输入Prediction为+...。\xe2\x84\xb9 组 1 中发生错误：理想 = 0。

\n

\n

我怎样才能以这种方式变异，我想引用前面的行变异？（除非它是第一行，在这种情况下使用 InitialPrice）

\n

[编辑]根据评论者，我尝试了累积，这是一个我不太熟悉的函数：

\n

mydiamonds.testdata <- mydiamonds.testdata %>% \n  mutate(\n    Prediction = accumulate(.f = function(.) {\n      \n    .$CoefIntercept + \n      \n      # here's the hard bit. If it's the first row in the group, use InitialPrice, else use the value of the previous prediction\n      (.$CoefLogLagCumPrice * ifelse(.$rn == 1, .$InitialPrice, lag(.$Prediction))) + \n      \n      (.$CoefLogRn * log(.$rn)) + \n      (.$CoefIdeal * .$Ideal)\n      \n      }))\nError: Problem with `mutate()` input `Prediction`.\nx argument ".x" is missing, with no default\n\xe2\x84\xb9 Input `Prediction` is `accumulate(...)`.\n\xe2\x84\xb9 The error occurred in group 1: Ideal = 0.\n

\n

Answer 1

正如您所说，您不习惯这个相当复杂的功能，这里有一些解释。

purrr::accumulate()用于计算逐行递归运算。它的第一个参数.x是您想要累积的变量。它的第二个参数.f是一个应该有 2 个参数的函数：当前结果cur和下一个评估值val。第一次.f被调用时，cur等于.x[1]（默认情况下），然后它等于之前返回的结果.f。

purrr::accumulate2()允许我们使用第二个变量.y进行迭代。第一个值.y总是被忽略，因为.f此时已经知道要返回什么。因此，.y应该比 短一项.x。

不幸的是，只有accumulate()和accumulate2()是您需要accumulate3()或paccumulate()积累的 rn、理想和价格。

但是，通过使用row_number()和cur_data()，您可以欺骗accumulate2()以按照您的意愿行事：

CoefIntercept = coeffs['(Intercept)']
CoefLogLagCumPrice = coeffs['log(lag(CumPrice))']
CoefLogRn = coeffs['log(rn)']
CoefIdeal = coeffs['Ideal']

mydiamonds.testdata <- mydiamonds %>% 
  ungroup() %>% 
  select(-CumPrice) %>% 
  mutate(
    Prediction = accumulate2(.x=InitialPrice, .y=row_number()[-1], 
                             .f=function(acc, nxt, row) {
      db=cur_data_all()
      rn = db$rn[row]
      Ideal = db$Ideal[row]
      CoefIntercept +
        (CoefLogLagCumPrice * acc) +
        (CoefLogRn * log(rn)) +
        (CoefIdeal * Ideal)
      
    }) %>% unlist()
  )
mydiamonds.testdata

# A tibble: 53,940 x 4
#     Ideal    rn InitialPrice Prediction
#     <dbl> <int>        <int>      <dbl>
# 1       0     1          326       326 
# 2       0     2          326       322.
# 3       0     3          326       318.
# 4       0     4          326       313.
# 5       0     5          326       309.
# 6       0     6          326       305.
# 7       0     7          326       301.
# 8       0     8          326       297.
# 9       0     9          326       294.
# 10      0    10          326       290.

编辑：还有另一种更清晰的使用.init参数的方法，因为InitialPrice除了第一个值之外，该列从未真正使用过。这允许直接使用参数，但它不适用于具有更多协变量的更复杂模型。

mydiamonds.testdata <- mydiamonds %>% 
  ungroup() %>% 
  select(-CumPrice) %>% 
  mutate(
    Prediction = accumulate2(.x=Ideal[-1], .y=rn[-1], 
                             .init=InitialPrice[1],
                             .f=function(rslt, Ideal, rn) {
      CoefIntercept +
        (CoefLogLagCumPrice * rslt) +
        (CoefLogRn * log(rn)) +
        (CoefIdeal * Ideal)
      
    }) %>% unlist()
  )