fuzziness argument in fuzzy_control function

[lucapizzimbone]

Hi,
I would like to highlight that using the tsclust function with the following configuration:

tsclust(data, k = 3,
distance = "dtw",
seed = 123,
type = "fuzzy",
centroid = "fcmdd",
contol = fuzzy_control(
fuzziness = 2,
iter.max = 100L
))

when changing the fuzziness parameter, there is no change in the membership degrees of the data. In particular, for a fuzziness = 1.000001, I would expect having for each data row one membership degree close to 1 and the others close to zero. However, any value of fuzziness > 1 provide the same results in the membership degrees.

Thank you in advance for the attention..
Grateful,
Luca

[asardaes]

Hello, can you elaborate on why you have that expectation? For example, if we assume that the truth is that there's only 1 cluster for the whole data set, trying to attempt fuzzy clustering with k > 1 would likely yield a membership degree of 1 / k for each entity in the data regardless of the other parameters, no?

[lucapizzimbone]

Dear Alexis, thank you for the swift response. my observation refers to the following statement (Fuzzy clustering - Wikipedia <https://en.wikipedia.org/wiki/Fuzzy_clustering>) [image: image.png] on the above basis, with a parameter m (fuzzifier in the dtwcluster library), I would probably expect membership values equal to 1 or 0, rather than float numbers. However, I noticed that the fuzziness parameter has only influence when: centroid = "fcm" when using centroid = "fcmdd" any there is no influence of the fuzziness value. Perhaps is the correct behaviour but from my side I cannot say muchmore at the moment. I will investigate further from my side too, and let you know. Thank you in advance for the help. Grateful, Luca I

…

On Wed, Nov 20, 2024 at 7:24 PM Alexis Sardá ***@***.***> wrote: Hello, can you elaborate on why you have that expectation? For example, if we assume that the truth is that there's only 1 cluster for the whole data set, trying to attempt fuzzy clustering with k > 1 would likely yield a membership degree of 1 / k for each entity in the data regardless of the other parameters, no? — Reply to this email directly, view it on GitHub <#79 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ATTE7DISEKQVKHFVRT7TIGL2BTHU5AVCNFSM6AAAAABSEYQAKCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIOBZGI3TIOJQHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[asardaes]

Hi Luca, github didn't render your email correctly, so I'm not sure which statement you were referring to, but if it's from the example, I don't think that's generalizable, that example concretely has most values closer to the extremes, and even then they're "massaging" the membership matrix into a single dimension in a way that could be confusing when compared to other cases.

[lucapizzimbone]

Hi Alexis, Thank you for the follow up. What I am trying to say is that when using centroid = "fcmdd" (see below set up of the tsclust function), for any valid value of fuzziness parameter, you will get always the same membership degree results, which seems to be incorrect. This is valid with any number of cluster and fuzziness value. Could you eventually advice which part of the source code is calculating the membership degree when "fcmdd" is selected as centroid. mvc <- tsclust(mydata, k = 5, distance = c("dtw"), seed = 555, type = "fuzzy", centroid = "fcmdd", preproc = NULL, contol = fuzzy_control( fuzziness = 2, # <----- any value here give always the same results when centroid = "fcmdd", rather than "fcm" iter.max = 100L ) Thank you in advance, Grateful, Luca

…

On Fri, Nov 22, 2024 at 10:17 PM Alexis Sardá ***@***.***> wrote: Hi Luca, github didn't render your email correctly, so I'm not sure which statement you were referring to, but if it's from the example, I don't think that's generalizable, that example concretely has most values closer to the extremes, and even then they're "massaging" the membership matrix into a single dimension in a way that could be confusing when compared to other cases. — Reply to this email directly, view it on GitHub <#79 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ATTE7DNH3UGOS2HASSQNTDT2B6NNPAVCNFSM6AAAAABSEYQAKCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIOJUHAZDMNJZGI> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[asardaes]

The logic for partitional and fuzzy clustering is more or less shared, the calculation you care about should happen here:

dtwclust/R/CLUSTERING-partitional-fuzzy.R

Line 44 in 9efcdb0

cluster <- family@cluster(distmat = distmat, m = control$fuzziness)

You should be able to debug the code and see the values at each step with:

debug(dtwclust:::pfclust)
# your tsclust call here