w В Пнд, 31/03/2008 в 01:40 +0200, Paco Pinto пишет: > Hi, > > I've read some papers concerning the duration modeling (in special > Yoshimura et al. 1998) and I understand the great advantages of using > a decision-trees. However it seems that the questions at each node are > different depending on the previous nodes.... probably a bit of > pseudo-code can help me to explain: > > (the IFs will represent the nodes) > > IF A THEN > IF B THEN > (...) > ELSE > IF C THEN > (...) > > >From the papers questions B and C are different, so if we have A a > set of questions will be made, if we have NOT A a different set of > questions will be made. > > Shouldn't all the questions be made in all the data subspaces ? I > mean, question B can be important if we have NOT A. > > Could you share with my some opinions on this ? > Many thanks, Well, you can use many data structures in machine learning. Decision tree is only one of the possible methods. About your question, yeah, it's sensible to join some leafs of the tree to compress it. IF A THEN goto check_c IF b then goto check_c check C: if C THEN and this was actually done for say LTS tasks in flite: http://www.cs.cmu.edu/~awb/papers/ISCA01/flite/node7.html Practically it just compresses the tree, complicating the training process. It won't give you better accuracy for the obvious reason. So there is no much sense to use WFST instead of usual tree except when you look for very small size.
Attachment:
signature.asc
Description: =?koi8-r?q?=FC=D4=C1?==?koi8-r?q?_=DE=C1=D3=D4=D8?= =?koi8-r?q?_=D3=CF=CF=C2=DD=C5=CE=C9=D1?= =?koi8-r?q?_=D0=CF=C4=D0=C9=D3=C1=CE=C1?= =?koi8-r?q?_=C3=C9=C6=D2=CF=D7=CF=CA?= =?koi8-r?q?_=D0=CF=C4=D0=C9=D3=D8=C0?=