best way to apply a list of functions to a dataset ?

Hi:

This might be a little easier (?):

library(datasets)
skewness <- function(x) mean(scale(x)^3)
mean.gt.med <- function(x) mean(x)>median(x)

# ------
# construct the function to apply to each variable in the data frame
f <- function(x) c(mean = mean(x), sd = sd(x), skewness = skewness(x),
???? median = median(x), mean.gt.med = mean.gt.med(x))

# map function to each variable with lapply and combine with do.call():
do.call(rbind, lapply(attitude, f))
?????????????? mean??????? sd??? skewness median mean.gt.med
rating???? 64.63333 12.172562 -0.35792491?? 65.5?????????? 0
complaints 66.60000 13.314757 -0.21541749?? 65.0?????????? 1
privileges 53.13333 12.235430? 0.37912287?? 51.5?????????? 1
learning?? 56.36667 11.737013 -0.05403354?? 56.5?????????? 0
raises???? 64.63333 10.397226? 0.19754317?? 63.5?????????? 1
critical?? 74.76667? 9.894908 -0.86577893?? 77.5?????????? 0
advance??? 42.93333 10.288706? 0.85039799?? 41.0?????????? 1

HTH,
Dennis


On Mon, Jul 19, 2010 at 10:51 PM, Glen Barnett <glnbrntt at gmail.com> wrote:

Assuming I have a matrix of data (or under some restrictions that will
become obvious, possibly a data frame), I want to be able to apply a
list of functions (initially producing a single number from a vector)
to the data and produce a data frame (for compact output) with column
1 being the function results for the first function, column 2 being
the results for the second function and so on - with each row being
the columns of the original data.

The obvious application of this is to produce summaries of data sets
(a bit like summary() does on numeric matrices), but with user
supplied functions. I am content for the moment to leave it to the
user to supply functions that work with the data they supply so as to
produce results that will actually be data-frame-able, though I'd like
to ultimately make it a bit nicer than it currently is without
compromising the niceness of the output in the "good" cases.

The example below is a simplistic approach to this problem (it should
run as is). I have named it "fapply" for fairly obvious reasons, but
added the ".1" because it doesn't accept multidimensional arrays. I
have included the output I generated, which is what I want. There are
some obvious generalizations (e.g. being able to include functions
like range(), say, that produce several values on a vector, rather
than one, making the user's life simpler when a function already does
most of what they need).

The question is: this looks like a silly approach, growing a list
inside a for loop. Also I recall reading that if you find yourself
using "do.call" you should probably be doing something else.

So my question: Is there a better way to implement a function like this?

Or, even better, is there already a function that does this?

## example function and code to apply a list of functions to a matrix
(here a numeric data frame)

library(datasets)

fapply.1 <- function(x, fun.l, colnames=fun.l){
out.l <- list() ? # starts with an empty list
for (i in seq_along(fun.l)) out.l[[i]] <- apply(x,2,fun.l[[i]]) ? #
loop through list of functions

# set up names and make into a data frame
names(out.l) <- colnames
attr(out.l,"row.names") <- names(out.l[[1]])
attr(out.l,"class") <- "data.frame"
out.l
}

skewness <- function(x) mean(scale(x)^3) ? ? ?#define a simple numeric
function
mean.gt.med <- function(x) mean(x)>median(x) ?# define a simple
non-numeric fn
flist <- c("mean","sd","skewness","median","mean.gt.med") # make list
of fns to apply

fapply.1(attitude,flist)
? ? ? ? ? ? ? mean ? ? ? ?sd ? ?skewness median mean.gt.med
rating ? ? 64.63333 12.172562 -0.35792491 ? 65.5 ? ? ? FALSE
complaints 66.60000 13.314757 -0.21541749 ? 65.0 ? ? ? ?TRUE
privileges 53.13333 12.235430 ?0.37912287 ? 51.5 ? ? ? ?TRUE
learning ? 56.36667 11.737013 -0.05403354 ? 56.5 ? ? ? FALSE
raises ? ? 64.63333 10.397226 ?0.19754317 ? 63.5 ? ? ? ?TRUE
critical ? 74.76667 ?9.894908 -0.86577893 ? 77.5 ? ? ? FALSE
advance ? ?42.93333 10.288706 ?0.85039799 ? 41.0 ? ? ? ?TRUE

## end code and output

So did I miss something obvious?

Any suggestions as far as style or simple stability-enhancing
improvements would be handy.

regards,
Glen

Hi Dennis,

Thanks for the reply.

Yes, that's easier, but the conversion to a matrix with rbind has
converted the output of that final function to a numeric.

I included that last function in the example secifically to preclude
people assuming that functions would always return the same type.

I guess this doesn't matter too much for a logical, but what if
instead the function returned a character (say "mean", "median", or
"equal" - indicating which one was larger, or "equal" which could
easily happen with discrete data). This precludes using rbind (which I
also used at first, before I noticed that sometimes I could have
functions that don't return numerics).

Hi:

On Tue, Jul 20, 2010 at 5:37 PM, Glen Barnett<glnbrntt at gmail.com>  wrote:

Hi Dennis,

Thanks for the reply.

Yes, that's easier, but the conversion to a matrix with rbind has
converted the output of that final function to a numeric.

If you look at the output of lapply(attitude, f), you'll see that the
conversion from logical to numeric has already taken place. Different
components of lists can have different types, but within a component, all of
the elements must have the same class.
You can patch up the result as follows:

x<- data.frame(do.call(rbind, lapply(attitude, f)))
x[, 5]<- as.logical(x[, 5])

x

                mean        sd    skewness median mean.gt.med
rating     64.63333 12.172562 -0.35792491   65.5       FALSE
complaints 66.60000 13.314757 -0.21541749   65.0        TRUE
privileges 53.13333 12.235430  0.37912287   51.5        TRUE
learning   56.36667 11.737013 -0.05403354   56.5       FALSE
raises     64.63333 10.397226  0.19754317   63.5        TRUE
critical   74.76667  9.894908 -0.86577893   77.5       FALSE
advance    42.93333 10.288706  0.85039799   41.0        TRUE

but if you're doing this sort of thing over a large data frame such fixes
may be impractical.


I included that last function in the example secifically to preclude

people assuming that functions would always return the same type.

There is a plyr solution, although it's a little more long-winded than I'd
prefer in the end:

library(ggplot2)
# melt the data frame so that the variables become factor levels
  ma<- melt(attitude)
Using  as id variables

dim(ma)

[1] 210   2
# Use ddply to get the set of summaries by variable:

ddply(ma, .(variable), summarise, mean = mean(value), sd = sd(value),
        skewness = skewness(value), median = median(value),
        mean.gt.med = mean.gt.med(value))
     variable     mean        sd    skewness median mean.gt.med
1     rating 64.63333 12.172562 -0.35792491   65.5       FALSE
2 complaints 66.60000 13.314757 -0.21541749   65.0        TRUE
3 privileges 53.13333 12.235430  0.37912287   51.5        TRUE
4   learning 56.36667 11.737013 -0.05403354   56.5       FALSE
5     raises 64.63333 10.397226  0.19754317   63.5        TRUE
6   critical 74.76667  9.894908 -0.86577893   77.5       FALSE
7    advance 42.93333 10.288706  0.85039799   41.0        TRUE

Notice that now the logical class of mean.gt.med is preserved.  The trick
with ddply() in package plyr is that, in this case, it is convenient to melt
the data frame first before doing the summarizations. This is because
ddply() requires a grouping variable - in this example, the groups are the
variables themselves.

HTH,
Dennis

I guess this doesn't matter too much for a logical, but what if

instead the function returned a character (say "mean", "median", or
"equal" - indicating which one was larger, or "equal" which could
easily happen with discrete data). This precludes using rbind (which I
also used at first, before I noticed that sometimes I could have
functions that don't return numerics).

Glen


On Tue, Jul 20, 2010 at 6:55 PM, Dennis Murphy<djmuser at gmail.com>  wrote:

Hi:

This might be a little easier (?):

library(datasets)
skewness<- function(x) mean(scale(x)^3)
mean.gt.med<- function(x) mean(x)>median(x)

# ------
# construct the function to apply to each variable in the data frame
f<- function(x) c(mean = mean(x), sd = sd(x), skewness = skewness(x),
      median = median(x), mean.gt.med = mean.gt.med(x))

# map function to each variable with lapply and combine with do.call():
do.call(rbind, lapply(attitude, f))
                mean        sd    skewness median mean.gt.med
rating     64.63333 12.172562 -0.35792491   65.5           0
complaints 66.60000 13.314757 -0.21541749   65.0           1
privileges 53.13333 12.235430  0.37912287   51.5           1
learning   56.36667 11.737013 -0.05403354   56.5           0
raises     64.63333 10.397226  0.19754317   63.5           1
critical   74.76667  9.894908 -0.86577893   77.5           0
advance    42.93333 10.288706  0.85039799   41.0           1

HTH,
Dennis


On Mon, Jul 19, 2010 at 10:51 PM, Glen Barnett<glnbrntt at gmail.com>

wrote:

Assuming I have a matrix of data (or under some restrictions that will
become obvious, possibly a data frame), I want to be able to apply a
list of functions (initially producing a single number from a vector)
to the data and produce a data frame (for compact output) with column
1 being the function results for the first function, column 2 being
the results for the second function and so on - with each row being
the columns of the original data.

The obvious application of this is to produce summaries of data sets
(a bit like summary() does on numeric matrices), but with user
supplied functions. I am content for the moment to leave it to the
user to supply functions that work with the data they supply so as to
produce results that will actually be data-frame-able, though I'd like
to ultimately make it a bit nicer than it currently is without
compromising the niceness of the output in the "good" cases.

The example below is a simplistic approach to this problem (it should
run as is). I have named it "fapply" for fairly obvious reasons, but
added the ".1" because it doesn't accept multidimensional arrays. I
have included the output I generated, which is what I want. There are
some obvious generalizations (e.g. being able to include functions
like range(), say, that produce several values on a vector, rather
than one, making the user's life simpler when a function already does
most of what they need).

The question is: this looks like a silly approach, growing a list
inside a for loop. Also I recall reading that if you find yourself
using "do.call" you should probably be doing something else.

So my question: Is there a better way to implement a function like this?

Or, even better, is there already a function that does this?

## example function and code to apply a list of functions to a matrix
(here a numeric data frame)

library(datasets)

fapply.1<- function(x, fun.l, colnames=fun.l){
out.l<- list()   # starts with an empty list
for (i in seq_along(fun.l)) out.l[[i]]<- apply(x,2,fun.l[[i]])   #
loop through list of functions

# set up names and make into a data frame
names(out.l)<- colnames
attr(out.l,"row.names")<- names(out.l[[1]])
attr(out.l,"class")<- "data.frame"
out.l
}

skewness<- function(x) mean(scale(x)^3)      #define a simple numeric
function
mean.gt.med<- function(x) mean(x)>median(x)  # define a simple
non-numeric fn
flist<- c("mean","sd","skewness","median","mean.gt.med") # make list
of fns to apply

fapply.1(attitude,flist)
               mean        sd    skewness median mean.gt.med
rating     64.63333 12.172562 -0.35792491   65.5       FALSE
complaints 66.60000 13.314757 -0.21541749   65.0        TRUE
privileges 53.13333 12.235430  0.37912287   51.5        TRUE
learning   56.36667 11.737013 -0.05403354   56.5       FALSE
raises     64.63333 10.397226  0.19754317   63.5        TRUE
critical   74.76667  9.894908 -0.86577893   77.5       FALSE
advance    42.93333 10.288706  0.85039799   41.0        TRUE

## end code and output

So did I miss something obvious?

Any suggestions as far as style or simple stability-enhancing
improvements would be handy.

regards,
Glen

ddply(ma, .(variable), summarise, mean = mean(value), sd = sd(value),
? ? ? skewness = skewness(value), median = median(value),
? ? ? mean.gt.med = mean.gt.med(value))