Basandomi sulla libSVM, ho iniziato lo sviluppo del software per SVC.

Devo dire che la disponibilità dei vari ricercatori nel fornire il codice da loro usato per effettuare i test è stata abbastanza scarna, a partire da Vapnik et al. e passando tra tutti coloro che negli ultimi anni hanno lavorato per eliminare il collo di bottiglia di questo approccio (cluster labeling).

Soltanto gli autori di

• bibtex | link
  J. Lee and D. Lee, "An Improved Cluster Labeling Method for Support Vector Clustering," IEEE Trans. Pattern Anal. Mach. Intell., vol. 27, iss. 3, pp. 461-464, 2005.
  @article{svcimproved, Address = {Washington, DC, USA},
  Author = {Jaewook Lee and Daewon Lee},
  Date-Added = {2007-04-28 18:28:29 +0200},
  Date-Modified = {2007-06-19 15:19:40 +0200},
  Doi = {http://dx.doi.org/10.1109/TPAMI.2005.47},
  Issn = {0162-8828},
  Journal = {IEEE Trans. Pattern Anal. Mach. Intell.},
  Keywords = {clustering, SVM},
  Note = {Member-Jaewook Lee},
  Number = {3},
  Pages = {461–464},
  Publisher = {IEEE Computer Society},
  Title = {An Improved Cluster Labeling Method for Support Vector Clustering},
  Url = {http://doi.ieeecomputersociety.org/10.1109/TPAMI.2005.47},
  Volume = {27},
  Year = {2005},
  Bdsk-File-1 = {YnBsaXN0MDDUAQIDBAUGBwpZJGFyY2hpdmVyWCR2ZXJzaW9uVCR0b3BYJG9iamVjdHNfEA9OU0tleWVkQXJjaGl2ZXISAAGGoNEICVRyb290gAGoCwwXGBkaHiVVJG51bGzTDQ4PEBMWWk5TLm9iamVjdHNXTlMua2V5c1YkY2xhc3OiERKABIAFohQVgAKAA4AHXHJlbGF0aXZlUGF0aFlhbGlhc0RhdGFfEFkuLi8uLi8uLi9QYXBlcnMvTGVlL0FuIEltcHJvdmVkIENsdXN0ZXIgTGFiZWxpbmcgTWV0aG9kIGZvciBTdXBwb3J0IFZlY3RvciBDbHVzdGVyaW5nLnBkZtIbDxwdV05TLmRhdGFPEQJYAAAAAAJYAAIAAAlEb2N1bWVudHMAAAAAAAAAAAAAAAAAAAAAAAC+zniuSCsAAAA3JUEfQW4gSW1wcm92ZWQgQ2×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×1c3RlciBMYWJlbGluZyBNZXRob2QgZm9yIFN1cHBvcnQgVmVjdG9yIENsdXN0ZXJpbmcucGRmAAATABIvVm9sdW1lcy9Eb2N1bWVudHMAFQACABf//wAAgAbSHyAhIlgkY2xhc3Nlc1okY2xhc3NuYW1loyIjJF1OU011dGFibGVEYXRhVk5TRGF0YVhOU09iamVjdNIfICYnoickXE5TRGljdGlvbmFyeQAIABEAGwAkACkAMgBEAEkATABRAFMAXABiAGkAdAB8AIMAhgCIAIoAjQCPAJEAkwCgAKoBBgELARMDbwNxA3YDfwOKA44DnAOjA6wDsQO0AAAAAAAAAgEAAAAAAAAAKAAAAAAAAAAAAAAAAAAAA8E=},
  Bdsk-Url-1 = {http://doi.ieeecomputersociety.org/10.1109/TPAMI.2005.47},
  Bdsk-Url-2 = {http://dx.doi.org/10.1109/TPAMI.2005.47}
}

mi hanno gentilmente risposto inviandomi il codice (putroppo Matlab) da loro usato per i test.

Sto perciò sviluppando questo software (che avrei dovuto fare comunque) con una visione più ampia in mente. Per ora la struttura è essenziale e fortemente legata a libSVM:

Classi fondamentali

• una classe astratta AbstractTrainer, che fornisce la fornisce l’interfaccia per l’implementazione della parte di training per il Support Vector Clustering.
• una classe AbstractLabeler, che fornisce la fornisce l’interfaccia per l’implementazione della parte di cluster labeling per il SVC

Prime concretizzazioni

• una classe OneClassTrainer, che implementa il training per il SVC utilizzando la One-class classification della libSVM
• [da terminare] una classe CGLabeler, che implementa nel modo più banale (e inefficiente) l’algoritmo cluster labeling proposto in
  • bibtex | link
    A. Ben-Hur, D. Horn, H. T. Siegelmann, and V. Vapnik, "Support Vector Clustering," Journal of Machine Learning Research, vol. 2, pp. 125-137, 2001.
    @article{svc,
  author = {A. Ben-Hur and D. Horn and H. T. Siegelmann and V. Vapnik},
  Date-Modified = {2007-06-19 14:44:40 +0200},
  Journal = {Journal of Machine Learning Research},
  Keywords = {clustering, SVM, gaussian kernel},
  Pages = {125-137},
  Title = {Support Vector Clustering},
  Url = {http://citeseer.ist.psu.edu/hur01support.html},
  Volume = 2, Year = 2001, Bdsk-File-1 = {YnBsaXN0MDDUAQIDBAUGBwpZJGFyY2hpdmVyWCR2ZXJzaW9uVCR0b3BYJG9iamVjdHNfEA9OU0tleWVkQXJjaGl2ZXISAAGGoNEICVRyb290gAGoCwwXGBkaHiVVJG51bGzTDQ4PEBMWWk5TLm9iamVjdHNXTlMua2V5c1YkY2xhc3OiERKABIAFohQVgAKAA4AHXHJlbGF0aXZlUGF0aFlhbGlhc0RhdGFfEDUuLi8uLi8uLi9QYXBlcnMvQmVuLUh1ci9TdXBwb3J0IFZlY3RvciBDbHVzdGVyaW5nLnBkZtIbDxwdV05TLmRhdGFPEQHqAAAAAAHqAAIAAAlEb2N1bWVudHMAAAAAAAAAAAAAAAAAAAAAAAC+zniuSCsAAAA3IEUdU3VwcG9ydCBWZWN0b3IgQ2×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},
  Bdsk-Url-1 = {http://citeseer.ist.psu.edu/hur01support.html}
}

Questa struttura permette di implementare agevolmente le diverse versioni di cluster labeling proposte in letteratura e al contempo di sviluppare diversi trainer, compreso quello per la formulazione Least Squares delle SVM, magari implementando l’algoritmo proposto in

• bibtex | link
  S. S. Keerthi and S. K. Shevade, "SMO algorithm for Least Squares SVM formulations," Neural Computation, vol. 15, pp. 487-507, 2003.
  @article{smo4lssvm03,
  author = {S.S. Keerthi and S.K. Shevade},
  Date-Added = {2007-05-19 19:11:38 +0200},
  Date-Modified = {2007-06-19 14:48:12 +0200},
  Journal = {Neural Computation},
  Keywords = {svm, ls-svm},
  Month = {February},
  Pages = {487–507},
  Title = {SMO algorithm for Least Squares SVM formulations},
  Url = {http://ieeexplore.ieee.org/iel5/8672/27485/01223730.pdf},
  Volume = {15},
  Year = {2003},
  Bdsk-File-1 = {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},
  Bdsk-Url-1 = {http://ieeexplore.ieee.org/iel5/8672/27485/01223730.pdf}
}