[datatable-help] merge/join/match

Gabor Grothendieck ggrothendieck at gmail.com
Sat May 4 11:46:10 CEST 2013


One further comment on nomatch=0 weirdness.  It seems that the value
of nomatch= is the row index of the row of X to return if a row in Y
matches no row in X here: X[Y,,nomatch=?]   In ordinary R indexing
using an index value of 0 means drop the corresponding component and
NA means return an NA.  nomatch=1 would presumably return the first
row of X for non-matching rows of Y but, in fact, nomatch= seems to be
restricted to 0 and NA as any other value generates an error message
to this effect. Likely it was decided that values other than 0 and NA
would be too bizarre and most likely represent user error.   If any.y=
were used then it would naturally be logical and this artificial
distinction (i.e .between 0/NA on one hand and everything else on the
other hand) would not have to be made.

On Fri, May 3, 2013 at 6:41 PM, Gabor Grothendieck
<ggrothendieck at gmail.com> wrote:
> In thinking about this a bit more I can see the argument for leaving
> the default at nomatch=NA. Consider these examples of indexing:
>
>> letters[27]
> [1] NA
>> BOD[7,]
>    Time demand
> NA   NA     NA
>
> nomatch=NA seems more compatible with these examples than nomatch=0.
>
> (At the same time this does not mean we could not also change the
> argument name from nomatch= to all.y= and add the other merge
> arguments (all.x=, by.x=, by.y=, by=) as well since it remains the
> case that R's merge() seems closer than R's match() to this
> functionality regardless of the default.)
>
>
> On Fri, May 3, 2013 at 4:42 PM, Gabor Grothendieck
> <ggrothendieck at gmail.com> wrote:
>> One can view data.table's generalization of indexing as the
>> realization that all indexing can conceptually be viewed as merging
>> where indexing with numeric values corresponds to merging with the
>> data.table's row numbers and indexing with logical values, L, is
>> equivalent to merging with which(L) so there are really not two types:
>> indexing and merging but just one type: merging that covers them all.
>>
>>
>> On Fri, May 3, 2013 at 1:01 PM, Arunkumar Srinivasan
>> <aragorn168b at gmail.com> wrote:
>>> I am wondering if performing X[Y] as a "merge" in correspondence with R's
>>> base "merge", if the basic idea of "i" becomes confusing. That is, when "i"
>>> is not a data.table in X[i] it indexes by rows. When `i` is a data.table,
>>> instead of the current definition which is in par with the subletting
>>> operation that use `i` (here data.table) as an index to subset X and then
>>> JOIN both X and Y, we say, here X and Y are data.tables and we perform a
>>> merge. I think this becomes confusing regarding the purpose of `i`.
>>>
>>> Remember that the main purpose of having the X[Y] is to have the flexibility
>>> of using `j` to to filter/subset only the desired columns. So, for example
>>> if you want to get 1 column of Y out of 100 columns when joining, you do:
>>> X[Y, list(cols_of_x, one_col_of_y)] and here, it doesn't go with the
>>> traditional definition of merge.
>>>
>>> As much as I like the idea of having consistent syntax, I also love the
>>> feature of X[Y, j]. So I'm confused as to how to deal with this.
>>>
>>> Arun
>>>
>>> On Friday, May 3, 2013 at 6:54 PM, Gabor Grothendieck wrote:
>>>
>>> I think that from the viewpoint of compatibility and convenience it
>>> would be best to implement all.x and all.y and not rely on swapping X
>>> and Y. SQLite did something like this (they implemented left join but
>>> not right join based on the idea that all you have to do is swap join
>>> arguments) but the problem with it is that it adds a layer of mental
>>> specification effort if the actual problem is better stated in the
>>> unsupported orientation.
>>>
>>> On Fri, May 3, 2013 at 12:49 PM, Eduard Antonyan
>>> <eduard.antonyan at gmail.com> wrote:
>>>
>>> Arun, it only needs the addition of smth like X[Y, keep.all = TRUE], all of
>>> the other merge options already exist as either X[Y] or Y[X] with or without
>>> nomatch = 0/NA.
>>>
>>>
>>> On Fri, May 3, 2013 at 11:45 AM, Arunkumar Srinivasan
>>> <aragorn168b at gmail.com> wrote:
>>>
>>>
>>> Gabor,
>>>
>>> Very true. I suppose your request is that the x[i] where `i` is a
>>> data.table should have the same set of options like R's base `merge`
>>> function, like, by.y=TRUE, by.x=TRUE or all=TRUE. I like the idea by itself.
>>> However, I am not able to think of a way to do this. I mean, I find the
>>> syntax X[Y, by.x=TRUE] weird / not making sense. That is, to me even though
>>>
>>> X[Y] is equal to Y[X, by.y=TRUE] (or) X[Y, by.x=TRUE] (ignoring the
>>> reordered columns) the latter 2 don't seem to make sense/is redundant (maybe
>>> it's because I am used to this syntax).
>>>
>>> Arun
>>>
>>> On Friday, May 3, 2013 at 5:57 PM, Gabor Grothendieck wrote:
>>>
>>> In my last post it should have read:
>>>
>>> That X[Y] is not the same as Y[X] is analogous to the fact that
>>> merge(X, Y, all.y=TRUE) is not the same as merge(Y, X, all.y=TRUE)
>>>
>>> On Fri, May 3, 2013 at 11:55 AM, Gabor Grothendieck
>>> <ggrothendieck at gmail.com> wrote:
>>>
>>> Assuming same-named keys, then these are all the same except possibly
>>> for row and column order:
>>>
>>> X[Y,,nomatch=0]
>>> Y[X,,nomatch=0]
>>> merge(X, Y)
>>> merge(Y, X)
>>>
>>> That X[Y] is not the same as Y[X] is analogous to the fact that
>>> merge(X, Y, all.x=TRUE) is not the same as merge(Y, X, all.x=TRUE)
>>>
>>> On Fri, May 3, 2013 at 11:46 AM, Arunkumar Srinivasan
>>> <aragorn168b at gmail.com> wrote:
>>>
>>> Gabor,
>>>
>>> X[Y] and Y[X] are not necessarily the same operations (meaning, they don't
>>> *have* to give the same output). However, merge(X,Y) and merge(Y,X) *have*
>>> to provide the same output (except for the column order and names). In
>>> that
>>> sense, a join is a bit different from a merge, no?
>>>
>>> Arun
>>>
>>> On Friday, May 3, 2013 at 5:36 PM, Gabor Grothendieck wrote:
>>>
>>> Yes, except that is not really what happens since match() only matches
>>> one row whereas with mult="all", the default, all rows are matched
>>> which is not really matching in the sense of match(). The current
>>> naming confuses matching with joining and its really the latter that
>>> is being done.
>>>
>>> Regarding the existence of merge the advantage of [ is that it will
>>> automatically only take the columns needed so merge is not really
>>> equivalent to [ in all respects. Furthermore having to use different
>>> constructs for different types of merge seems awkward.
>>>
>>>
>>> On Fri, May 3, 2013 at 11:27 AM, Eduard Antonyan
>>> <eduard.antonyan at gmail.com> wrote:
>>>
>>> Btw the way I think about the "nomatch" name is as follows - normally X[Y]
>>> tries to match rows of Y with rows of X, and then "nomatch" tells it what
>>> to
>>> do when there is *no match*.
>>>
>>>
>>> On Fri, May 3, 2013 at 10:23 AM, Eduard Antonyan
>>> <eduard.antonyan at gmail.com>
>>> wrote:
>>>
>>>
>>> To clarify - that behavior is already implemented in merge (more
>>> specifically merge.data.table). I don't really have a view on having it in
>>> X[Y] as well - I don't like all.x and all.y as the names, since there are
>>> no
>>> params named 'x' and 'y' in [.data.table (as opposed to merge), but some
>>> param that would do a full outer join could certainly be added.
>>>
>>>
>>> On Fri, May 3, 2013 at 10:09 AM, Gabor Grothendieck
>>> <ggrothendieck at gmail.com> wrote:
>>>
>>>
>>> Yes, sorry. Its nomatch= which presumably derives from the parameter
>>> of the same name in the match() function. If the idea of the nomatch=
>>> name was to leverage off existing argument names in R then I would
>>> prefer all.y= to be consistent with merge() in place of nomatch= since
>>> we are really merging/joining rather than just matching. That would
>>> also allow extension to all types of join by adding all.an x= argument
>>> too.
>>>
>>> On Fri, May 3, 2013 at 10:59 AM, Eduard Antonyan
>>> <eduard.antonyan at gmail.com> wrote:
>>>
>>> I would prefer nomatch=0 as a default though, simply because that's
>>> what I
>>> do most of the time :)
>>>
>>>
>>> On Fri, May 3, 2013 at 9:57 AM, Eduard Antonyan
>>> <eduard.antonyan at gmail.com>
>>> wrote:
>>>
>>>
>>> A correction - the param is called "nomatch", not "match".
>>>
>>> This use case seems like smth a user shouldn't really do - in an ideal
>>> world you should have them both keyed by the same-name column.
>>>
>>> As is, my view on it is that data.table is correcting the user mistake
>>> of
>>> naming the column in Y - y, instead of x, and so the output makes
>>> sense and
>>> I don't see the need of complicating the behavior by adding more cases
>>> one
>>> has to go through to figure out what the output columns would be.
>>> Similar to
>>> asking for X[J(c("b", "c", "d"))] - you wouldn't want an anonymous
>>> column
>>> there, would you?
>>>
>>>
>>>
>>> On Fri, May 3, 2013 at 6:18 AM, Gabor Grothendieck
>>> <ggrothendieck at gmail.com> wrote:
>>>
>>>
>>> I am moving this discussion which started with mdowle to the list.
>>>
>>> Consider this example slightly modified from the data.table FAQ:
>>>
>>> X = data.table(x=c("a","a","b","b","b","c","c"), foo=1:7, key="x")
>>> Y = data.table(y=c("b","c","d"), bar=c(4,2,3))
>>> out <- X[Y]; out
>>>
>>> x foo bar
>>> 1: b 3 4
>>> 2: b 4 4
>>> 3: b 5 4
>>> 4: c 6 2
>>> 5: c 7 2
>>> 6: d NA 3
>>>
>>> Note that the first column of the output is labelled x even though
>>> the
>>> data to produce it comes from y, e.g. "d" in out$x is not in X$x but
>>> does appear in Y$y so clearly the data is coming from y as opposed to
>>> x . In terms of SQL the above would be written:
>>>
>>> select Y.y as x, ...
>>>
>>> and the need to renamne the first column of out suggests that there
>>> may be a deeper problem here.
>>>
>>> Here are some ideas to address this (they would require changes to
>>> data.table):
>>>
>>> - the default of X[Y,, match=NA] would be changed to a default of
>>> X[Y,,match=0] so that it corresponds to the defaults in R's merge and
>>> in SQL joins.
>>>
>>> - the column name of the first column in the example above would be
>>> changed to y if match=0 but be left at x if match=NA. In the case
>>> that match=0 (the proposed new default) x and y are equal so the
>>> first
>>> column can be validly labelled as x but in the case that match=NA
>>> they
>>> are not so y would be used as the column name.
>>>
>>> - the name match= does seem a bit misleading since R's match only
>>> matches one item in the target whereas in data.table match matches
>>> many if mult="all" and that is the default. Perhaps some thought
>>> should be given to a name change here?
>>>
>>> The above would seem to correspond more closely to R's merge and SQL
>>> join defaults. Any use cases or other comments?
>>>
>>> --
>>> Statistics & Software Consulting
>>> GKX Group, GKX Associates Inc.
>>> tel: 1-877-GKX-GROUP
>>> email: ggrothendieck at gmail.com
>>> _______________________________________________
>>> 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
>>>
>>>
>>>
>>>
>>> --
>>> Statistics & Software Consulting
>>> GKX Group, GKX Associates Inc.
>>> tel: 1-877-GKX-GROUP
>>> email: ggrothendieck at gmail.com
>>>
>>>
>>>
>>>
>>> --
>>> Statistics & Software Consulting
>>> GKX Group, GKX Associates Inc.
>>> tel: 1-877-GKX-GROUP
>>> email: ggrothendieck at gmail.com
>>> _______________________________________________
>>> 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
>>>
>>>
>>>
>>>
>>> --
>>> Statistics & Software Consulting
>>> GKX Group, GKX Associates Inc.
>>> tel: 1-877-GKX-GROUP
>>> email: ggrothendieck at gmail.com
>>>
>>>
>>>
>>>
>>> --
>>> Statistics & Software Consulting
>>> GKX Group, GKX Associates Inc.
>>> tel: 1-877-GKX-GROUP
>>> email: ggrothendieck at gmail.com
>>>
>>>
>>>
>>>
>>> --
>>> Statistics & Software Consulting
>>> GKX Group, GKX Associates Inc.
>>> tel: 1-877-GKX-GROUP
>>> email: ggrothendieck at gmail.com
>>>
>>>
>>
>>
>>
>> --
>> Statistics & Software Consulting
>> GKX Group, GKX Associates Inc.
>> tel: 1-877-GKX-GROUP
>> email: ggrothendieck at gmail.com
>
>
>
> --
> Statistics & Software Consulting
> GKX Group, GKX Associates Inc.
> tel: 1-877-GKX-GROUP
> email: ggrothendieck at gmail.com



-- 
Statistics & Software Consulting
GKX Group, GKX Associates Inc.
tel: 1-877-GKX-GROUP
email: ggrothendieck at gmail.com


More information about the datatable-help mailing list