在处理NaN时,data.table"list"与":="

注意到data.table的一些奇怪的行为,希望有人比我能解释的更了解data.table.

说我有这个data.table:

library(data.table)
DT <- data.table(
  C1 = c(rep("A", 4), rep("B",4), rep("C", 4)),
  C2 = c(rep("a", 3), rep("b",3), rep("c",3), rep("d",3)),
  Val = c(1:5, NaN, NaN, 8,9,10,NaN,12))

DT
    C1 C2 Val
 1:  A  a   1
 2:  A  a   2
 3:  A  a   3
 4:  A  b   4
 5:  B  b   5
 6:  B  b NaN
 7:  B  c NaN
 8:  B  c   8
 9:  C  c   9
10:  C  d  10
11:  C  d NaN
12:  C  d  12

现在,在我看来,以下两种方法应该生成相同的结果,但它们不会.

TEST1 <- DT[, agg := min(Val, na.rm = TRUE), by = c('C1', 'C2')]
TEST1 <- data.table(unique(TEST1[, c('C1','C2','agg'), with = FALSE]))

TEST2 <- DT[, list(agg = min(Val, na.rm = TRUE)), by = c('C1', 'C2')]

TEST1
   C1 C2 agg
1:  A  a   1
2:  A  b   4
3:  B  b   5
4:  B  c   8
5:  C  c   9
6:  C  d  10


TEST2
   C1 C2 agg
1:  A  a   1
2:  A  b   4
3:  B  b   5
4:  B  c NaN
5:  C  c   9
6:  C  d  10

如您所见,使用":="会为(C1 = B,C2 = c)生成最小值8.而list命令会生成NaN.有趣的是,对于(C1 = B,C2 = b)和(C1 = C,C2 = d),它们也有NaNs,list命令确实产生一个值.我相信这是因为在NaN首先在给定C1 C2组合的值之前的情况下,NaN结果.而在另外两个例子中,NaN来自一个值.

为什么会这样？

我注意到如果用NA替换NaN,则生成的值没有问题.

修复了这个问题,#1461刚刚开发,v1.9.7,提交2080.

require(data.table) # v1.9.7, commit 2080+
DT <- data.table(
     C1 = c(rep("A", 4), rep("B",4), rep("C", 4)),
     C2 = c(rep("a", 3), rep("b",3), rep("c",3), rep("d",3)),
     Val = c(1:5, NaN, NaN, 8,9,10,NaN,12))

DT[, list(agg = min(Val, na.rm = TRUE)), by = c('C1', 'C2')]
#    C1 C2 agg
# 1:  A  a   1
# 2:  A  b   4
# 3:  B  b   5
# 4:  B  c   8
# 5:  C  c   9
# 6:  C  d  10