[R] Why CLARA clustering method does not give the same classes as when I do clustering manually?

[Martin Maechler]

>>>>> David L Carlson <dcarlson at tamu.edu>
>>>>>     on Sun, 21 Feb 2016 16:55:56 +0000 writes:

    > I do not think this is quite true. When the medoids are
    > not specified, pam/clara looks for a good initial set
    > (build phase) and then finds a local minimum of the
    > objective function (swap phase). Both pam/clara and kmeans
    > can find local minima that are not the global minimum. 

Indeed, thank you for the explanation so far.

    > If the build phase involves any random element, two runs
    > could produce different results. 

One of the important features of pam (over kmeans) is indeed
that the build phase is entirely deterministic and typically
much better than random starts.
But there is an option to pam() to skip tbe build phase and
start from user specified medoids, so you can also do random
starts as kmeans, by using
      pam(..., medoids = sample(n, k))

And now you should not throw  clara  and  pam together:
Clara does randomly choose a subset from the LARge data
Application ("LARA") and then runs the build phase (and more) with only
that subset; hence clara necessarily has an element of
randomness to it.
Though --- for historical reasons -- the default chosen by CLARA's
original authors had been to always chose the same random seed
by default (and a cheap non-R RNG).  For that reasons, for a
long time now, clara()  has got an argument 'rngR' which you
"should" set to TRUE in order to get the variability of random starts.

.. more feature requests are welcome!

Martin Maechler,
ETH Zurich          == maintainer("cluster")


    > If not, then the original
    > order of the data determines the final result, but the
    > final result is not necessarily the best one possible
    > (assuming the order of the data is irrelevant to the
    > analysis so we are not looking at observations taken along
    > a line in time or space). That is why kmeans includes an
    > argument to run the algorithm multiple times and pick the
    > best result.

    > -------------------------------------
    > David L Carlson Department of Anthropology Texas A&M
    > University College Station, TX 77840-4352

    > -----Original Message----- From: R-help
    > [mailto:r-help-bounces at r-project.org] On Behalf Of Sarah
    > Goslee Sent: Friday, February 19, 2016 1:47 PM To:
    > ABABAEI, Behnam Cc: r-help at r-project.org Subject: Re: [R]
    > Why CLARA clustering method does not give the same classes
    > as when I do clustering manually?

    > clara() is a version of pam() adapted to use large
    > datasets.

    > pam() uses the entire dataset, and should give results
    > identical to your manual procedure, or nearly so. clara()
    > works on subsets of the data, so it may give a slightly
    > different result each time you run it.

    > The default parameters for clara() are very small, so you
    > can get substantially different results from run to run on
    > a large dataset if you don't change them.

    > Sarah

    > On Fri, Feb 19, 2016 at 6:30 AM, ABABAEI, Behnam
    > <Behnam.ABABAEI at limagrain.com> wrote:
    >> Hi,
    >> 
    >> 
    >> I am using CLARA (in 'cluster' package). This method is
    >> supposed to assign each observation to the closest
    >> 'medoid'. But when I calculate the distance of medoids
    >> and observations manually and assign them manually, the
    >> results are slightly different (1-2 percent of occurrence
    >> probability). Does anyone know how clara calculates
    >> dissimilarities and why I get different clustering
    >> results?
    >> 
    >> 
    >> Behnam.

    > ______________________________________________
    > R-help at r-project.org mailing list -- To UNSUBSCRIBE and
    > more, see https://stat.ethz.ch/mailman/listinfo/r-help
    > PLEASE do read the posting guide
    > http://www.R-project.org/posting-guide.html and provide
    > commented, minimal, self-contained, reproducible code.

    > ______________________________________________
    > R-help at r-project.org mailing list -- To UNSUBSCRIBE and
    > more, see https://stat.ethz.ch/mailman/listinfo/r-help
    > PLEASE do read the posting guide
    > http://www.R-project.org/posting-guide.html and provide
    > commented, minimal, self-contained, reproducible code.