[datatable-help] Summing over many variables

Matthew Dowle mdowle at mdowle.plus.com
Fri Dec 31 15:13:12 CET 2010


Hi,

I don't quite follow all of that e.g. I'm thinking secondary keys at
points (manual now, or 'built-in' feature request). Anyway, sounds like
it's working. On point 3 I'm not sure that's really data.table, rather
the difference between a vectorized sum and apply()-ed sum. You should
see the same difference with a data.frame.

Where it isn't working (point 5) was there an error message or was the
result incorrect? Might be a clue that reveals a bug.

Matthew

On Wed, 2010-12-29 at 12:57 -0500, Joseph Voelkel wrote:
> Thanks, Matthew.
> 
> 1. Yes, you have the subsetting of j on your faq.
> 2. The double eval appears to handle this subsetting. In my smaller problem, which I am working on first, I have about 55K records and 250 variables. Using either A1+A2+A3+A4+A5 or eval(eval()) takes about 0.22 sec of user time. So, from this indirect measure, the two are equally efficient.
> 3. By the way, for this example, I used 6 keys, which happened for this problem to correspond to 1 record for each unique key. That is, my output also contained 55K records. I originally solved the problem by using apply with sum on the five columns that contained the A1 through A5 value, e.g. apply(DT1[,11:15,with=FALSE],1,sum). This operation took about 0.62 sec. So, even here, data.table is 3x faster than apply.
> 4. Of course, no key is really needed here, so if I just want to return the sum along with the key vars, I can just use DT1[,list(key1,key2,key3,key4,key5,key6,sum=A1+A2+A3+A4+A5)] which runs in under 0.01 seconds.
> 5. Finally, I tried your idea of removing the quote and just trying the one eval(). It worked with a simple contrived example, but not for my more complex one--I have no idea why not, because the two seem analogous...
> 
>  
> 
> -----Original Message-----
> From: Matthew Dowle [mailto:mdowlenoreply at virginmedia.com] On Behalf Of Matthew Dowle
> Sent: Tuesday, December 28, 2010 12:22 PM
> To: Joseph Voelkel
> Cc: datatable-help at lists.r-forge.r-project.org
> Subject: Re: [datatable-help] Summing over many variables
> 
> Glad that works. Thanks for posting back. One thintg with that approach
> is that data.table inspects the j expression to see which columns it
> uses. It only subsets the ones that are used, for efficiency. There's a
> faq on that I think. If the expression is wrapped up inside an eval I
> think it still inspects the j but I can't quite remember. I'd be
> surprised if that works with the double eval like that. If A runs from 1
> to 100 in your real data and you're taking many sub-sums of 5, then this
> could make a big difference. Try timing sum(A1) vs sum(A2+A3+A4+A5) with
> and without the eval(eval()). That should reveal whether the j is being
> inspected ok.
> Also looking at it again, you shouldn't need the quote() inside the text
> passed to parse. Then it's just a single eval and j inspection should be
> ok I think i.e. DT1[,eval(ASumExpr),by=grp] rather than
> DT1[,eval(eval(ASumExpr)),by=grp]
> 
> Matthew
> 
> 
> On Mon, 2010-12-27 at 13:23 -0500, Joseph Voelkel wrote:
> > I like Matthew's idea of flattening tables. But, as usual, I did not tell the whole story in my first post. I will probably want to look at many expressions, for example,
> > 
> > sum(A1+A2+A3+A4+A5)
> > sum(A2+A3+A4+A5+A6)
> > sum(A3+A4+A5+A6+A7)
> > sum((A1+A2)/2 - (A3+A4)/2)
> > 
> > To be able to investigate a sequence of these easily, I found (after some trial and error, and then thinking about it a bit more to try to make my problem look like one from the datatable-faq) that this will do the trick:
> > 
> > library(data.table)
> > 
> > # create data table
> > DT1<-data.table(A1=1:1000000,A2=1:1000000,A3=1:1000000,A4=1:1000000,A5=1:1000000,grp=rep(1:50000,each=20))
> > setkey(DT1,grp)
> > 
> > # Say I want DT1[,sum(A1+A2+A3+A4+A5),by=grp]
> > 
> > # First, create expression of interest, and convert it to data-table-useful form
> > ASumExpr<-parse(text=paste("quote(sum(",paste("A",1:5,sep="",collapse="+"),"))",sep=""))
> > # (Next few lines: to help me and maybe you see what this looks like...)
> > ASumExpr
> > str(ASumExpr)
> > eval(ASumExpr)
> > str(eval(ASumExpr))
> > str(quote(mean(x))) # from example in datatable-faq.pdf. So eval(ASumExpr) looks good
> > 
> > # long-hand typing method. OK for one or two, but not in general
> > system.time(dt2a<-DT1[,sum(A1+A2+A3+A4+A5),by=grp])
> > # formula method. This will be useful.
> > system.time(dt2b<-DT1[,eval(eval(ASumExpr)),by=grp])
> > 
> > identical(dt2a, dt2b)
> > 
> > # Fast and easy to write. Just what I wanted. Thanks again for the ideas that lead to this useful solution.
> > 
> > Joe V.
> > 
> > -----Original Message-----
> > From: Matthew Dowle [mailto:mdowlenoreply at virginmedia.com] On Behalf Of Matthew Dowle
> > Sent: Thursday, December 23, 2010 4:33 PM
> > To: Joseph Voelkel
> > Cc: datatable-help at lists.r-forge.r-project.org
> > Subject: Re: [datatable-help] Summing over many variables
> > 
> > 
> > Yes that's one way. We aren't that happy with using lapply in j as it
> > loses the benefit of data.table.
> > 
> > I tend to 'flatten' tables like this. Try to have few columns. In this
> > case it would be either a 3 column table (grp,colname,value) or maybe a
> > 4 column table if you ever want to group by "A" or
> > "B" (grp,letter,number,value). The query would then be
> > DT[,sum(value),by=list(grp,letter,number)]. You can then do pattern
> > matches and filters etc in the i rather than in the j e.g.
> > DT[letter=="A",sum(value),by=group] for just the "A"s. The answer comes
> > out in 'flat' format but you can always 'unflatten' the result to make
> > it look pretty or easier to read.  [Note that I sinned by using '==' in
> > the i just then invoking vector scan, so to avoid that for speed you
> > would setkey(letter,group) then DT["A",sum(value),by=group]], or getting
> > fancy if you only wanted some groups (say 1 and 3) then 'by without by'
> > e.g. DT[list("A",c(1,3)),sum(value)].
> > 
> > 'flat' is a common way to use data.table to store higher dimensional
> > data, and especially sparse higher dimensional data.
> > 
> > The 'grp.1' repetition is a problem I'd like to remove. It's related to
> > this feature request (but is almost a bug). At the moment you have to
> > remove the grp.1 afterwards.
> > https://r-forge.r-project.org/tracker/index.php?func=detail&aid=978&group_id=240&atid=978
> > 
> > Matthew
> > 
> > _______________________________________________
> > datatable-help mailing list
> > datatable-help at lists.r-forge.r-project.org
> > https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/datatable-help
> 
> 




More information about the datatable-help mailing list