Author:
Jinbo Bi bij2@rpi.edu
Contact:
Jinbo Bi
Department of Mathematical Sciences
Rensselaer Polytechnic Institute
110 8th Street
Troy, NY 12180
Goal:
To compare three support vector regression algorithms,
the first is our proposed RH-SVR method, see related references
for details, the second is the nu-SVR method, which is proposed
by Bernhard Scholkopf et.al., the third one is the classic
SVR method, see Vapnik's Statistical Learning Theory.
Source Codes:
1. cplexcpp.h -- header file for CPLEX interface
2. cplexcpp.C -- cplexcpp class
3. sv_crqp.C -- implement RH-SVR algorithm
4. sv_nuqp.C -- implement nu-SVR algorithm
5. sv_rqp.C -- implement classic SVR algorithm
Other Files:
1. Makefile (for Unix Sun Solaris OS)
2. toydata (a sample toydata ready to run)
3. sv_crqp.par (a sample parameter file to feed in program SV_crqp)
4. sv_nuqp.par (a sample parameter file to feed in program SV_nuqp)
5. sv_rqp.par (a sample parameter file to feed in program SV_rqp)
Implemented in C++
It can be compiled on SUN Solaris OS and IBM AIX OS
if you have CPLEX software installed in the corresponding
system. You need to chance the Makefile in order to include
correct library location.
Compile:
1. make all
will generate all three programs, SV_crqp (RH-SVR method), SV_nuqp
(nu-SVR method), and SV_rqp (classic SVR method)
2. make sv_crqp
will only generate the SV_crqp executable program
3. make sv_nuqp
will only generate the SV_nuqp executable program
4. make sv_rqp
will generate the SV_rqp executable program
5. make cc
will clean up those .o files
6. make clean
will clean up all executable files and .o files
Help:
After you successfully compile the source codes, you can type
in the program name such as SV_crqp without any arguments, and
you can see a brief help.
Run:
Three files have exactly the same format for use. For example,
using the following command
SV_crqp toydata sv_crqp.par 20 0.3
means to perform 20 bootstraps with 30% of total examples in the
data file "toydata" left out as test data in each bootstrap. The
choices of the parameters such as C, epsilon and others are
provided in sv_crqp.par.