当需要考虑机器公差时，按数字列合并 data.tables

Question

Many have seen the issue with using == to compare to floating point numbers. == fails to return TRUE but all.equal works.

x <- sqrt(2)
x^2 == 2
#> [1] FALSE
all.equal(x^2, 2)
#> [1] TRUE

My issue comes from the need to join to data.table s by a numeric column where == will fail to find the matching pairs.

I have considered coercing the numeric values to characters, but that option has too many other potiential errors. I have considered rounding the values, but that to, in the application I need, will create more problems.

Here is simple example of a join that is failing because DT1$x == DT2$x will return FALSE when it would be preferable to have the return be TRUE .

library(data.table)
packageVersion("data.table")
#> [1] '1.12.8'

DT1 <- data.table(x = sqrt(1:10), v1 = 1:10)
DT2 <- data.table(x = 1:10, v2 = LETTERS[1:10])

# set x to its square
DT1[, x := x^2]

# left join
merge(DT1, DT2, by = "x", all.x = TRUE)
#>      x v1   v2
#>  1:  1  1    A
#>  2:  2  2 <NA>
#>  3:  3  3 <NA>
#>  4:  4  4    D
#>  5:  5  5 <NA>
#>  6:  6  6 <NA>
#>  7:  7  7 <NA>
#>  8:  8  8 <NA>
#>  9:  9  9    I
#> 10: 10 10 <NA>

How can I specify a left join by a numeric column key such that the machine tolerance in the comparison is accounted for? ^{Created on 2020-04-06 by the reprex package (v0.3.0)}

Answer 1

You could use roll = "nearest" . Note that only the last column specified in on = can be rolling.

library(data.table)
DT1[DT2,on = "x", roll = "nearest"]
    x v1 v2
 1:  1  1  A
 2:  2  2  B
 3:  3  3  C
 4:  4  4  D
 5:  5  5  E
 6:  6  6  F
 7:  7  7  G
 8:  8  8  H
 9:  9  9  I
10: 10 10  J

I suspect the problem is more complicated than this simple case, but you could subsequently filter joins that do not meet a certain threshold of difference.

Data

DT1 <- data.table(x = sqrt(1:10), v1 = 1:10)
DT2 <- data.table(x = 1:10, v2 = LETTERS[1:10])
DT1[, x := x^2]