data.table interpolando linearmente os valores de NA sem grupos

Eu queria preencher alguns valores de NA em uma tabela de dados sem grupos. Por favor, considere este extrato de data.table representando tempo e distâncias:

library(data.table)
df <- data.frame(time = seq(7173, 7195, 1), dist = c(31091.33, NA, 31100.00, 31103.27, NA, NA, NA, NA, 31124.98, NA,31132.81, NA, NA, NA, NA, 31154.19, NA, 31161.47, NA, NA, NA, NA, 31182.97))
DT<- data.table(df)

Eu quero na tabela de dados DT, preencher os valores de NA com uma função dependendo do valor não-NA antes / depois. Como exemplo, escrevendo uma função em j para substituir cada instrução

DT[2, dist := (31091.33 + (31100-31091.33) / 2)]

então

DT[5:8, dist := (31103.27 + "something" * (31124.98 - 31103.27) / 5)]

etc ...

O código é explicado em linha. Você pode excluir as colunas temporárias usando df[,dist_before := NULL], por exemplo.

library(data.table)
df=data.table(time=seq(7173,7195,1),dist=c(31091.33,NA,31100.00,31103.27,NA,NA,NA,
NA,31124.98,NA,31132.81,NA,NA,NA,NA,31154.19,NA,31161.47,NA,NA,NA,NA,31182.97))
df
#>     time     dist
#>  1: 7173 31091.33
#>  2: 7174       NA
#>  3: 7175 31100.00
#>  4: 7176 31103.27
#>  5: 7177       NA
#>  6: 7178       NA
#>  7: 7179       NA
#>  8: 7180       NA
#>  9: 7181 31124.98
#> 10: 7182       NA
#> 11: 7183 31132.81
#> 12: 7184       NA
#> 13: 7185       NA
#> 14: 7186       NA
#> 15: 7187       NA
#> 16: 7188 31154.19
#> 17: 7189       NA
#> 18: 7190 31161.47
#> 19: 7191       NA
#> 20: 7192       NA
#> 21: 7193       NA
#> 22: 7194       NA
#> 23: 7195 31182.97
#>     time     dist
# Carry forward the last non-missing observation
df[,dist_before := nafill(dist, "locf")]
# Bring back the next non-missing dist
df[,dist_after := nafill(dist, "nocb")]
# rleid will create groups based on run-lengths of values within the data.
# This means 4 NA's in a row will be grouped together, for example.
# We then count the missings and add 1, because we want the 
# last NA before the next non-missing to be less than the non-missing value.
df[, rle := rleid(dist)][,missings := max(.N +  1 , 2), by = rle][]
#>     time     dist dist_before dist_after rle missings
#>  1: 7173 31091.33    31091.33   31091.33   1        2
#>  2: 7174       NA    31091.33   31100.00   2        2
#>  3: 7175 31100.00    31100.00   31100.00   3        2
#>  4: 7176 31103.27    31103.27   31103.27   4        2
#>  5: 7177       NA    31103.27   31124.98   5        5
#>  6: 7178       NA    31103.27   31124.98   5        5
#>  7: 7179       NA    31103.27   31124.98   5        5
#>  8: 7180       NA    31103.27   31124.98   5        5
#>  9: 7181 31124.98    31124.98   31124.98   6        2
#> 10: 7182       NA    31124.98   31132.81   7        2
#> 11: 7183 31132.81    31132.81   31132.81   8        2
#> 12: 7184       NA    31132.81   31154.19   9        5
#> 13: 7185       NA    31132.81   31154.19   9        5
#> 14: 7186       NA    31132.81   31154.19   9        5
#> 15: 7187       NA    31132.81   31154.19   9        5
#> 16: 7188 31154.19    31154.19   31154.19  10        2
#> 17: 7189       NA    31154.19   31161.47  11        2
#> 18: 7190 31161.47    31161.47   31161.47  12        2
#> 19: 7191       NA    31161.47   31182.97  13        5
#> 20: 7192       NA    31161.47   31182.97  13        5
#> 21: 7193       NA    31161.47   31182.97  13        5
#> 22: 7194       NA    31161.47   31182.97  13        5
#> 23: 7195 31182.97    31182.97   31182.97  14        2
#>     time     dist dist_before dist_after rle missings
# .SD[,.I] will get us the row number relative to the group it is in. 
# For example, row 5 dist is calculated as
# dist_before + 1 * (dist_after - dist_before)/5
df[is.na(dist), dist := dist_before + .SD[,.I] *
                     (dist_after - dist_before)/(missings), by = rle]
df[]
#>     time     dist dist_before dist_after rle missings
#>  1: 7173 31091.33    31091.33   31091.33   1        2
#>  2: 7174 31095.67    31091.33   31100.00   2        2
#>  3: 7175 31100.00    31100.00   31100.00   3        2
#>  4: 7176 31103.27    31103.27   31103.27   4        2
#>  5: 7177 31107.61    31103.27   31124.98   5        5
#>  6: 7178 31111.95    31103.27   31124.98   5        5
#>  7: 7179 31116.30    31103.27   31124.98   5        5
#>  8: 7180 31120.64    31103.27   31124.98   5        5
#>  9: 7181 31124.98    31124.98   31124.98   6        2
#> 10: 7182 31128.90    31124.98   31132.81   7        2
#> 11: 7183 31132.81    31132.81   31132.81   8        2
#> 12: 7184 31137.09    31132.81   31154.19   9        5
#> 13: 7185 31141.36    31132.81   31154.19   9        5
#> 14: 7186 31145.64    31132.81   31154.19   9        5
#> 15: 7187 31149.91    31132.81   31154.19   9        5
#> 16: 7188 31154.19    31154.19   31154.19  10        2
#> 17: 7189 31157.83    31154.19   31161.47  11        2
#> 18: 7190 31161.47    31161.47   31161.47  12        2
#> 19: 7191 31165.77    31161.47   31182.97  13        5
#> 20: 7192 31170.07    31161.47   31182.97  13        5
#> 21: 7193 31174.37    31161.47   31182.97  13        5
#> 22: 7194 31178.67    31161.47   31182.97  13        5
#> 23: 7195 31182.97    31182.97   31182.97  14        2
#>     time     dist dist_before dist_after rle missings

Você pode usar o approx função para fazer interpolação linear.

Para cada grupo de NAs, obtenha esse subconjunto de DTmais as linhas antes e depois. Em seguida, aplique approxa esse subconjunto do distvetor, com o nargumento approxigual ao número de linhas no subconjunto .N.

DT[, g := rleid(dist)]

DT[is.na(dist), dist := {
      i <- .I[c(1, .N)] + c(-1, 1)
      DT[i[1]:i[2], approx(dist, n = .N)$y[-c(1, .N)]]
  }, by = g]

Ou, sem approx

DT[, g := rleid(dist)]

DT[is.na(dist), dist := {
      i <- .I[c(1, .N)] + c(-1, 1)
      DT[i[1]:i[2], dist[1] + 1:(.N - 2)*(dist[.N] - dist[1])/(.N - 1)]
  }, by = g]

edit: desde que essa resposta foi aceita, devo salientar que outras respostas são mais rápidas e a segunda parte da resposta da @ dww é basicamente o meu primeiro bloco de código, mas com a parte desnecessária do agrupamento removida (por isso é mais simples e rápida).

2 outras opções:

1) junção de rolamento:

DT[is.na(dist), dist := {
        x0y0 <- DT[!is.na(dist)][.SD, on=.(time), roll=Inf, .(time=x.time, dist=x.dist)]
        x1y1 <- DT[!is.na(dist)][.SD, on=.(time), roll=-Inf, .(time=x.time, dist=x.dist)]
        (x1y1$dist - x0y0$dist) / (x1y1$time - x0y0$time) * (time - x0y0$time) + x0y0$dist
    }]
DT

2) outra variante próxima da resposta de Smingerson usando nafill

DT[, dist := {
    y0 <- nafill(dist, "locf")
    x0 <- nafill(replace(time, is.na(dist), NA), "locf")
    y1 <- nafill(dist, "nocb")
    x1 <- nafill(replace(time, is.na(dist), NA), "nocb")
    fifelse(is.na(dist), (y1 - y0) / (x1 - x0) * (time - x0) + y0, dist)
}]

código de temporização:

library(data.table)
set.seed(0L)
# df=data.frame(time=seq(7173,7195,1),dist=c(31091.33,NA,31100.00,31103.27,NA,NA,NA,NA,31124.98,NA,31132.81,NA,NA,NA,NA,31154.19,NA,31161.47,NA,NA,NA,NA,31182.97))
# DT=data.table(df)
nr <- 1e7
nNA <- nr/2
DT <- data.table(time=1:nr, dist=replace(rnorm(nr), sample(1:nr, nNA), NA_real_))

DT00 <- copy(DT)
DT01 <- copy(DT)
DT1 <- copy(DT)
DT20 <- copy(DT)
DT201 <- copy(DT)
DT202 <- copy(DT)
DT21 <- copy(DT)

mtd00 <- function() {
    DT00[, g := rleid(is.na(dist))]

    DT00[is.na(dist), dist := {
        i <- .I[c(1, .N)] + c(-1, 1)
        DT00[i[1]:i[2], approx(dist, n = .N)$y[-c(1, .N)]]
    }, by = g]
}

mtd01 <- function() {
    DT01[, g := rleid(is.na(dist))]

    DT01[is.na(dist), dist := {
        i <- .I[c(1, .N)] + c(-1, 1)
        DT01[i[1]:i[2], dist[1] + 1:(.N - 2)*(dist[.N] - dist[1])/(.N - 1)]
    }, by = g]
}

mtd1 <- function() {
    DT1[,dist_before := nafill(dist, "locf")]
    DT1[,dist_after := nafill(dist, "nocb")]
    DT1[, rle := rleid(dist)][,missings := max(.N +  1 , 2), by = rle][]
    DT1[is.na(dist), dist_before + .SD[,.I] *
            (dist_after - dist_before)/(missings), by = rle]
}


mtd20 <- function() {
    DT20[is.na(dist), {
        x0y0 <- DT20[!is.na(dist)][.SD, on=.(time), roll=Inf, .(time=x.time, dist=x.dist)]
        x1y1 <- DT20[!is.na(dist)][.SD, on=.(time), roll=-Inf, .(time=x.time, dist=x.dist)]
        (x1y1$dist - x0y0$dist) / (x1y1$time - x0y0$time) * (time - x0y0$time) + x0y0$dist
    }]
}

mtd201 <- function() {
    i <- DT201[, is.na(dist)]
    DT201[(i), {
        x0y0 <- DT201[(!i)][.SD, on=.(time), roll=Inf, .(time=x.time, dist=x.dist)]
        x1y1 <- DT201[(!i)][.SD, on=.(time), roll=-Inf, .(time=x.time, dist=x.dist)]
        (x1y1$dist - x0y0$dist) / (x1y1$time - x0y0$time) * (time - x0y0$time) + x0y0$dist
    }]
}

mtd202 <- function() {
    i <- DT201[is.na(dist), which=TRUE]
    DT201[i, {
        x0y0 <- DT201[-i][.SD, on=.(time), roll=Inf, .(time=x.time, dist=x.dist)]
        x1y1 <- DT201[-i][.SD, on=.(time), roll=-Inf, .(time=x.time, dist=x.dist)]
        (x1y1$dist - x0y0$dist) / (x1y1$time - x0y0$time) * (time - x0y0$time) + x0y0$dist
    }]
}


mtd21 <- function() {
    DT21[, {
        y0 <- nafill(dist, "locf")
        x0 <- nafill(replace(time, is.na(dist), NA), "locf")
        y1 <- nafill(dist, "nocb")
        x1 <- nafill(replace(time, is.na(dist), NA), "nocb")
        fifelse(is.na(dist), (y1 - y0) / (x1 - x0) * (time - x0) + y0, dist)
    }]
}

bench::mark(
    #mtd00(), mtd01(), 
    #mtd1(),
    mtd20(), mtd201(), mtd202(),
    mtd21(), check=FALSE)

horários:

# A tibble: 4 x 13
  expression      min   median `itr/sec` mem_alloc `gc/sec` n_itr  n_gc total_time result            memory            time    gc            
  <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl> <int> <dbl>   <bch:tm> <list>            <list>            <list>  <list>        
1 mtd20()       1.19s    1.19s     0.838    1.01GB    1.68      1     2      1.19s <dbl [5,000,000]> <df[,3] [292 x 3~ <bch:t~ <tibble [1 x ~
2 mtd201()      1.12s    1.12s     0.894  954.06MB    0.894     1     1      1.12s <dbl [5,000,000]> <df[,3] [341 x 3~ <bch:t~ <tibble [1 x ~
3 mtd202()      1.16s    1.16s     0.864  858.66MB    1.73      1     2      1.16s <dbl [5,000,000]> <df[,3] [392 x 3~ <bch:t~ <tibble [1 x ~
4 mtd21()    729.93ms 729.93ms     1.37   763.11MB    1.37      1     1   729.93ms <dbl [10,000,000~ <df[,3] [215 x 3~ <bch:t~ <tibble [1 x ~

edit: para abordar comentários sobre o uso is.na(dist)várias vezes:

set.seed(0L)
nr <- 1e7
nNA <- nr/2
DT <- data.table(time=1:nr, dist=replace(rnorm(nr), sample(1:nr, nNA), NA_real_))
DT20 <- copy(DT)
DT201 <- copy(DT)
DT202 <- copy(DT)

mtd20 <- function() {
    DT20[is.na(dist), dist := {
        x0y0 <- DT20[!is.na(dist)][.SD, on=.(time), roll=Inf, .(time=x.time, dist=x.dist)]
        x1y1 <- DT20[!is.na(dist)][.SD, on=.(time), roll=-Inf, .(time=x.time, dist=x.dist)]
        (x1y1$dist - x0y0$dist) / (x1y1$time - x0y0$time) * (time - x0y0$time) + x0y0$dist
    }]
}

mtd201 <- function() {
    i <- DT201[, is.na(dist)]
    DT201[(i), dist := {
        x0y0 <- DT201[(!i)][.SD, on=.(time), roll=Inf, .(time=x.time, dist=x.dist)]
        x1y1 <- DT201[(!i)][.SD, on=.(time), roll=-Inf, .(time=x.time, dist=x.dist)]
        (x1y1$dist - x0y0$dist) / (x1y1$time - x0y0$time) * (time - x0y0$time) + x0y0$dist
    }]
}

mtd202 <- function() {
    i <- DT201[is.na(dist), which=TRUE]
    DT201[i, dist := {
        x0y0 <- DT201[-i][.SD, on=.(time), roll=Inf, .(time=x.time, dist=x.dist)]
        x1y1 <- DT201[-i][.SD, on=.(time), roll=-Inf, .(time=x.time, dist=x.dist)]
        (x1y1$dist - x0y0$dist) / (x1y1$time - x0y0$time) * (time - x0y0$time) + x0y0$dist
    }]
}

horários:

# A tibble: 3 x 13
  expression      min   median `itr/sec` mem_alloc `gc/sec` n_itr  n_gc total_time result                    memory             time     gc               
  <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl> <int> <dbl>   <bch:tm> <list>                    <list>             <list>   <list>           
1 mtd20()      24.1ms   25.8ms      37.5    1.01GB    13.6     11     4      294ms <df[,2] [10,000,000 x 2]> <df[,3] [310 x 3]> <bch:tm> <tibble [15 x 3]>
2 mtd201()     24.8ms   25.6ms      38.2  954.07MB     8.19    14     3      366ms <df[,2] [10,000,000 x 2]> <df[,3] [398 x 3]> <bch:tm> <tibble [17 x 3]>
3 mtd202()       24ms   25.6ms      38.3   76.39MB     8.22    14     3      365ms <df[,2] [10,000,000 x 2]> <df[,3] [241 x 3]> <bch:tm> <tibble [17 x 3]>

Não vê muita diferença nos tempos quando reduz o número de is.na(dist)chamadas

Usando library(zoo)

DT[, dist := na.approx(dist)]

Como alternativa, se você preferir manter as funções R básicas em vez de usar outro pacote, poderá fazer

DT[, dist := approx(.I, dist, .I)$y]

Aqui está uma abordagem rcpp que percorre tudo de uma vez com passagem adicional para todos os elementos de NA.

Rcpp::sourceCpp(code = '
#include <Rcpp.h>
using namespace Rcpp;

// [[Rcpp::export]]
NumericVector rcpp_approx2D(IntegerVector x, NumericVector y) {
  double x_start = 0, y_start = 0, slope = 0;
  int count = 0;

  NumericVector y1 = clone(y); //added to not update-by-reference

  for(int i = 0; i < y1.size(); ++i){
    if (NumericVector::is_na(y1[i])){
      count++;
    } else {
      if (count != 0) {
        x_start = x[i-(count+1)];
        y_start = y1[i-(count+1)];
        slope = (y1[i] - y_start) / (x[i]- x_start);
        for (int j = 0; j < count; j++){
          y1[i-(count-j)] = y_start + slope * (x[i - (count - j)] - x_start);
        }
        count = 0;
      }
    }
  }
  return(y1);
}
')

Então em R:

DT[, rcpp_approx2D(time, dist)]