multi-labeled samples

18 views
Skip to first unread message

Ferhat KURTULMUŞ

unread,
Nov 26, 2013, 4:43:25 AM11/26/13
to bob-...@googlegroups.com
I don't know if it is a proper place to ask my question. I have data with multi-labeled samples. Namely, my data has instances which may belong to two independent classes in the same time (not multi-classes). My question raises the point that I want to perform one or more methods (eg., stepwise discriminant analysis) to eliminate ineffective features. And I wonder if the bob library supports multi-labeling for both stepdisc and training-testing with any classifier.

Thanks in advance,
Ferhat

Laurent El Shafey

unread,
Nov 26, 2013, 5:41:04 AM11/26/13
to bob-...@googlegroups.com
Hello,

This is definitely the good place to ask such a question.
AFAIK, we did not implement any specific multi-label classification
algorithm. To be honest, I'm not familiar with this topic and with how
exactly you formulate your problem (e.g. does all instance have exactly
two labels?). For instance, if you have n labels and each samples has
exactly two labels (among these 'n'), I guess that one option would be
to consider that this is a binomial(n,2)-class classification (one class
for each pair of labels) problem, and then to use a regular
binomial(n,2)-class classifier. The classification algorithms
implemented are listed in the user guide/tutorials of bob.

Cheers,
Laurent
> --
> -- You received this message because you are subscribed to the Google
> Groups bob-devel group. To post to this group, send email to
> bob-...@googlegroups.com. To unsubscribe from this group, send email
> to bob-devel+...@googlegroups.com. For more options, visit
> this group at https://groups.google.com/d/forum/bob-devel or directly
> the project website at http://idiap.github.com/bob/
> ---
> You received this message because you are subscribed to the Google
> Groups "bob-devel" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to bob-devel+...@googlegroups.com.
> For more options, visit https://groups.google.com/groups/opt_out.


Ferhat KURTULMUŞ

unread,
Nov 26, 2013, 7:09:36 AM11/26/13
to bob-...@googlegroups.com, laurent....@idiap.ch
My data is like:
                                                                                   labels     
      | feature 1 | feature 2 | feature 3 | .... feature n  | type 1 | type 2|
1                                                                            D          1
2                                                                            D           2
3                                                                            OD       1
4                                                                            OD       6
.                                                                             MD        1
.
.
n-samples

type 1 may be one of D, OD, MD. type 2 may be any number between 1 and 6.

26 Kasım 2013 Salı 12:41:04 UTC+2 tarihinde Laurent El Shafey yazdı:

Laurent El Shafey

unread,
Nov 27, 2013, 3:44:29 PM11/27/13
to bob-...@googlegroups.com
Hello,

We don't have any 'multi-label' algorithm implemented in bob.
But in your case, you could combine the two labels into a single label such as D1, D2, OD1, OD6, MD1 etc. You would hence have a single label with 18 possible values for it (instead of 2 labels with 3 and 6 possible values), and you could then use a 'regular' one label classifier. I don't know if there are better strategies for these kinds of problems, still.

Cheers,
Laurent

Ferhat KURTULMUŞ

unread,
Nov 29, 2013, 4:00:43 AM11/29/13
to bob-...@googlegroups.com, laurent....@idiap.ch
Thanks your response. It is what I had already tried first. I just wanted to try a different thing. Yet, I appreciate it for your help.

27 Kasım 2013 Çarşamba 22:44:29 UTC+2 tarihinde Laurent El-Shafey yazdı:
Reply all
Reply to author
Forward
0 new messages