data.tableroll=“nearest”返回多个结果-data.tableroll=“nearest”returnsmultipleresults

5  

Referring to Matt's answer there is an easy way to use all the 15 significant digits a double offers in order to properly select the closest matching row. Instead of working on the original values, one can scale the values up to ensure that the 15 significant digits lie above the 10^(-8) level. This could be done as follows:

参考Matt的答案,有一种简单的方法可以使用双重提供的所有15位有效数字,以便正确选择最接近的匹配行。可以将值向上扩展以确保15个有效数字位于10 ^( - 8)级别之上,而不是处理原始值。这可以如下完成:

orig_vals <- dt[,val]
scale_fact <- max(10^(trunc(log10(abs(orig_vals)))+8))
scaled_vals <- orig_vals * scale_fact
dt[,scaled_val:=scaled_vals]
setkey(dt,scaled_val)

Now, performing the rolling join

现在,执行滚动连接

scaled_x <- x*scale_fact
dt[J(scaled_x), roll="nearest"][, ifelse(is.na(val), NA_real_, rnk)]

# [1] 0.2727273

yields - as desired - a single value.

If also in the case of two identical key values only one row should be selected, the mult="first" argument can be added to the above data.table call.

产量 - 根据需要 - 单一价值。如果在两个相同的键值的情况下也只应选择一行,则可以将mult =“first”参数添加到上面的data.table调用中。

8  

Yes, data.table automatically applies a tolerance when joining and grouping numeric columns. The tolerance in v1.8.10 is sqrt(.Machine$double.eps) == 1.490116e-08. This comes directly from ?base::all.equal.

是的,data.table在连接和分组数字列时自动应用容差。 v1.8.10中的容差是sqrt(.Machine $ double.eps)== 1.490116e-08。这直接来自?base :: all.equal。

To illustrate, consider grouping :

为了说明,请考虑分组:

> dt
          rnk        val
1: 0.00000000 0.02337751
2: 0.09090909 0.02708315
3: 0.18181818 0.02750162
4: 0.27272727 0.02750162

> dt[,.N,by=val]
          val N
1: 0.02337751 1
2: 0.02708315 1
3: 0.02750162 2    # one group, size two
>

When you joined using dt[J(x), roll="nearest"], that x value matched to within tolerance and you got the group it matched to, as usual when a matching value occurs in a rolling join. roll="nearest" only applies to the values that don't match, outside tolerance.

当您使用dt [J(x),roll =“nearest”]加入时,该x值匹配在容差范围内,并且您获得与之匹配的组,就像通常在滚动连接中出现匹配值一样。 roll =“nearest”仅适用于不匹配的值,超出容差范围。

data.table considers the values in rows 3 and 4 of val to be equal. The thinking behind this is for convenience, since most of the time key values are really a fixed precision such as prices ($1.23) or recorded measurements to a specified precision (1.234567). We'd like to join and group such numerics even after multiplying them for example, without needing to code for machine accuracy ourselves. And we'd like to avoid confusion when numeric data displays as though it's equal in a table, but isn't due to very tiny differences in the bit representation.

data.table认为val的第3行和第4行中的值相等。这背后的想法是为了方便,因为大多数时候键值实际上是固定的精度,例如价格(1.23美元)或记录的测量到指定的精度(1.234567)。我们希望加入并组合这些数字,即使在它们相乘之后,也不需要自己编码机器精度。我们希望避免在数字数据显示时在表格中相等的混淆,但不是由于位表示的微小差异。

See ?unique.data.table for this example :

有关此示例,请参阅?unique.data.table:

DT = data.table(a=tan(pi*(1/4 + 1:10)), b=rep(1,10))   # example from ?all.equal
length(unique(DT$a))         # 10 strictly unique floating point values
all.equal(DT$a,rep(1,10))    # TRUE, all within tolerance of 1.0
DT[,which.min(a)]            # row 10, the strictly smallest floating point value
identical(unique(DT),DT[1])  # TRUE, stable within tolerance
identical(unique(DT),DT[10]) # FALSE

data.table is also stable within tolerance; i.e, when you group by a numeric, the original order of the items within that group are maintained as usual.

data.table在容忍范围内也是稳定的;即,当您按数字分组时,该组中项目的原始顺序将照常维护。

> dt$val[3]  dt[, row:=1:4]  # add a row number to illustrate
> dt[, list(.N, list(row)), by=val]
          val N  V2
1: 0.02337751 1   1
2: 0.02708315 1   2
3: 0.02750162 2 3,4
> dt[3:4, val:=rev(val)]   # swap the two values around
> dt$val[3] > dt$val[4]
[1] TRUE
> dt[, list(.N, list(row)), by=val]
          val N  V2
1: 0.02337751 1   1
2: 0.02708315 1   2
3: 0.02750162 2 3,4    # same result, consistent. stable within tolerance