Re: [R] "FANNY" function in R package "cluster"

From: Martin Maechler <maechler_at_stat.math.ethz.ch>
Date: Tue 31 May 2005 - 18:40:32 EST

>>>>> "Aamir" == Aamir M <intuitionist@gmail.com>
>>>>> on Mon, 30 May 2005 15:57:29 -0400 writes:

    Aamir> Dear All, I am attempting to use the FANNY fuzzy
    Aamir> clustering function in R (Kaufman & Rousseeuw, 1990),
    Aamir> found in the "cluster" package. I have run into a
    Aamir> variety of difficulties; the two most crucial
    Aamir> difficulties are enumerated below.

    Aamir> 1. Where is the 'm' parameter in FANNY?  In _Finding
    Aamir> Groups in Data: An Introduction to Cluster Analysis_
    Aamir> (1990) by Kaufman & Rousseeuw, the authors discuss
    Aamir> the FANNY algorithm. On pages 189-190, they attempt
    Aamir> to demonstrate an equivalence between Fuzzy c-Means     Aamir> (FCM) and FANNY.

The section is called "*Related* Methods and References" and they (as most Statisticians) use the word "k-Means"...

The first point in that section is to explain the NON-equivalence: fanny() works with arbitrary dissimilarities d[i,j] whereas all versions of k-means have to assume a euclidean measurement space. K&R show that *when* you have a data matrix X, and then use distances d[i,j] := || x_{.,i} - x_{.,j} ||^2 (i.e. SQUARED Euclidean distances), then fanny() does the same as fuzzy k-means. But they even say there, that they'd rather use the non-squared distance {as fanny() does by default}, i.e., exponent 1, not 2, for good robustness reasons. If this is your 'm' below, then fanny() already does what you might want by default. OTOH, you can always pass squared distances to fanny() ..

    Aamir> In doing so they, appear to be
    Aamir> assuming that the value of the 'm' parameter in FCM
    Aamir> (a measure of the fuzziness) is fixed at m=2.

there is no 'm' in the book there, but they talk about the exponent "^ 2" used in some places {but "^ 1" in other places}, notably in 5.2 "Why did we choose FANNY?"

There is no "fuzziness" parameter defined there, so can you be more specific?

    Aamir>  Although this is how FCM was originally
    Aamir> formulated, it eventually became apparent that m
    Aamir> should be a user-specified parameter and not a fixed
    Aamir> value of m=2.  My question, then, is
    Aamir> twofold. Firstly, am I correct in saying that Kaufman
    Aamir> & Rousseeuw have assumed m=2 in their formulation of
    Aamir> Fuzzy c-Means and FANNY? Secondly, is it possible to
    Aamir> modify the FANNY algorithm to allow
    Aamir> user-specification of the m (fuzziness) parameter?

Maybe, if you tell me what you are taking about. See the section 5.2 mentioned above
Is it the exponent 2 in u_{jv}^2 ?
That one is currently fixed at 2, and yes, that could be made a parameter though K & R warn against going all the way to "1" where their algorithm can happend to converge very slowly.

    Aamir> 2. What do I do with training data?  Is there any way
    Aamir> to use FANNY for assigning clustering membership
    Aamir> values to new, test data? In Fuzzy c-Means, new data
    Aamir> is compared to the cluster centers in order to assign
    Aamir> clustering membership values to the test
    Aamir> data. However, in FANNY these centers do not
    Aamir> exist.

correct. Note again that in general such centers don't even make sense, since the sample space may contain unordered categorical variables mixed with continuous ones {and you would be advised to use daisy(), not dist() to compute dissimilarities for such data}.

    Aamir> Is there, then, any way to compute the FANNY
    Aamir> clustering membership values of a test data point
    Aamir> without affecting the clustering membership values of
    Aamir> the training data? Perhaps there are enough
    Aamir> conditions to use the objective function as a way of
    Aamir> computing the membership values of the test data?

    Aamir> Aamir M University of Toronto

That's an interesting proposal, at least the way I choose to understand you :-)

Yes, why not look at the objective function C {eq.(1), p.182}

One could think of optimizing it with respect to new data only, by keeping all "old data" memberships.
For that to work, one would need the n dissimilarites

    d[i', j] where i' : `index for' new data

               j = 1,..,n : indices for training data. Is this feasible in your situation?

Alternatively, when we *did* assume ``all continuous'' data *and* the use of simple Euclidean distances, we could easily compute the cluster centers, determine (by minimization!) memberships for new observations.

In any case that needs some assumptions (and code!) currently not part of fanny().



R-help@stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html Received on Tue May 31 18:45:35 2005

This archive was generated by hypermail 2.1.8 : Fri 03 Mar 2006 - 03:32:16 EST