R KNN imputation - function returning erroneous results & missing help page

问题

I'm trying to impute some missing values in R using library(imputation) and kNNImpute(). The input data frame is 44 rows of 13 variables. There are 30 complete observations and 14 observations with missing values in 2 columns.

The code is saying it's imputing all the missing values; however, it's imputing the last 4 values as 0. From my reading of the code, this appears to be a flaw based on using 0 as a default for errors. My code:

# impute data
library(imputation)
knn_data <- kNNImpute(x, k= 5)

# examine kNNImpute code
kNNImpute

kNNImpute's code: See lines 4, 8 the function starting on line 24 and the 2nd line from the bottom (line 48):

[4]  prelim = impute.prelim(x)
[8]  x.missing = prelim$x.missing
[24] x.missing.imputed = t(apply(x.missing, 1, function(i) {...}
[48] x[missing.matrix2] = 0

??impute.prelim returns no results (the help page is missing). So, I can't examine this code.

However, the program flow for kNNImpute appears to be

[4]  # run a (seemingly undefined) screening function
[8]  # pull in the missing rows for later imputation
[24] # run imputation function
[48] # based on line [4] output, impute all "error rows"  == 0

Can anyone explain why this is happening and/or how to solve this problem?

FYI- I have emailed the package author a link to this page.

回答1:

Solution: I used code identical to the kNNImpute() function to impute the 4 improperly imputed values.

impute.fn <- function(scores, distances, raw_dist) {
  knn.values <- scores[c(as.integer(names(distances)))]
  knn.weights <- 1 - (distances / max(raw_dist))
  weighted.mean(knn.values, knn.weights)
}

# impute errors - rows 41-44 are improperly imputed 
  # rows 1-30 have non missing avlues
#---------------------------------------------------------
x.dist <- as.matrix(dist(x))
dist_41 <- x.dist[41, c(1:30)][order(x.dist[41, c(1:30)])]
...

# fix impute - column 1
x$ABC[41] <- impute.fn(x$ABC, dist_41[1:5], dist_41)
...

An appropriate answer from the package author (or other) would still be appreciated.

Note: I have re-written the imputation package for wKNN. Improved package can be found here: imputaton

来源：https://stackoverflow.com/questions/20294603/r-knn-imputation-function-returning-erroneous-results-missing-help-page

标签

knn