preProcess with knnImpute returns skewed non-missing values

[dianashams]

Hi,
I was testing preProcess imputations and noticed that when I apply the method to a data, it changes the original values even if those were not missing in "knnImpute" method, but the behaviour for "badImpute" is as expected, that is, it leaves known values as those are and imputes missing ones. Is this a bug? Does this affect downstream tasks when knnImpute is used? Thank you

Example code:
df = data.frame(matrix(rnorm(600, 0,1), nrow = 150, ncol = 4))
df[,1][sample(1:150, 30)] <- NA
df[,2][sample(1:150, 30)] <- NA
pp1 = preProcess(df, method = "bagImpute")
pp2 = preProcess(df, method = "knnImpute")
df1 = predict(pp1, df)
df2 = predict(pp2, df)
round(cbind("original" = df$X1, "bagImpute" = df1$X1, "knnImpute" = df2$X1),4)

 original bagImpute knnImpute

[1,] 0.5374 0.5374 __ 0.5743 __ => why knnImpute changes the values which were not missing to start with?..
[2,] 0.5118 0.5118 0.5449
[3,] -0.2418 -0.2418 -0.3195
[4,] -0.6785 -0.6785 -0.8204
[5,] 0.8837 0.8837 0.9716
[6,] 1.3987 1.3987 1.5622
[7,] -0.1245 -0.1245 -0.1849
[8,] 0.8306 0.8306 0.9106
[9,] -0.9723 -0.9723 -1.1575
[10,] NA 0.1627 -0.1757
[11,] 0.1003 0.1003 0.0729
[12,] -0.2509 -0.2509 -0.3300
[13,] 0.8210 0.8210 0.8996
[14,] -2.4520 -2.4520 -2.8547
[15,] NA 0.0507 -0.1487
[16,] NA -0.3666 -0.0224

[dianashams]

Please ignore - I understand that kNN's imputation needs to be de-normalised, this line of code

for (i in 1:ncol(df2)) {df2[,i] <- df2[,i] * pp2$std[i] + pp2$mean[i]}

makes all three columns as expected

round(cbind("original" = df$X1, "bagImpute" = df1$X1, "knnImpute" = df2$X1),4)