Campus Event Calendar
Event Entry
What and Who
Can support vector machine be a major classification method ? Part 2: Issues on SVM software design and the use of LIBSVM
Chih-Jen Lin
Taiwan
Talk
AG 1, AG 2, AG 3, AG 4
MPI Audience
Note: We use this to send email in the morning.
Date, Time and Location
Wednesday, 12 February 2003
15:00
-- Not specified --
46.1 - MPII
024
Saarbrücken
Abstract
Support vector machines (SVM) have been a promising classification
method. However, currently traditional methods such as decision trees
and Neural Networks remain the major tools used by practitioners
(e.g. the Kdnuggets Poll in 2002). It is important to investigate
obstacles left for transforming SVM from a hot machine learning topic
to a mainstream classification tool.
From users of our SVM software we realize that there is a huge gap
between sophisticated machine learning techniques and
practitioners. Therefore, users often improperly apply a
classification method or use only its primitive part. Then the result
(accuracy) is not satisfactory. Our past and future research are to
identify and study techniques which are easily enough to be adapted by
general SVM users. This will be the first part of the talk in which we
particularly focus on issues of SVM model selection. Different model
selection techniques will be presented and we explain which one might
be the most suitable for users.
method. However, currently traditional methods such as decision trees
and Neural Networks remain the major tools used by practitioners
(e.g. the Kdnuggets Poll in 2002). It is important to investigate
obstacles left for transforming SVM from a hot machine learning topic
to a mainstream classification tool.
From users of our SVM software we realize that there is a huge gap
between sophisticated machine learning techniques and
practitioners. Therefore, users often improperly apply a
classification method or use only its primitive part. Then the result
(accuracy) is not satisfactory. Our past and future research are to
identify and study techniques which are easily enough to be adapted by
general SVM users. This will be the first part of the talk in which we
particularly focus on issues of SVM model selection. Different model
selection techniques will be presented and we explain which one might
be the most suitable for users.
Contact
--email hidden
passcode not visible
logged in users only