file_format.tex

一种新颖的SVM算法

/*
 * File:        sample_list_file_format.txt
 * Purpose:     Description of the data file formats of the SVM
 *
 * Author:      Mark O. Stitson
 * Created:     00/00/97
 * Updated:     20/11/97
 *
 * 
 * Copyright (c) 1997  RHBNC London - All rights reserved
 * THIS IS PROPRIETARY SOURCE CODE of RHBNC London
 */

Format for loading and saving the sample_list_c:

The sample list is an ascii file only containing numbers. The first
few numbers indicate the exact format followed by the data.

The sample_list can load several formats but only saves one format.

The first number (int) of the sample_list file always contains the number of
examples in the file.

The second number (int) of the sample_list file always contains the
dimensionality of the input space.

The third number (int) of the sample list file determines the format
of the file. If it is positive, this is the dimensionality
classification of the examples (version -1), this has to be 1 in this
version. If it is negative it indicates the version number of the file
format.

Version -1:

The rest of the file simply consists of examples. First the input
values of an example then its classification.

Say we have four examples in two dimensional input space and the
classification follows the function f(x_1,x_2)=2 x x_1 + x_2.
The input file should look something like this:

4
2
1
1   1   3
1.5 3.4 6.4
1.2 0   2.4
0   3   3

Version 0:

The fourth number (int) indicates the dimensionality of the classification of
the examples. (In this version this has to be 1.)
The fifth number (0/1) indicates whether or not the data has been
pre-scaled. This is useful if other data should be scaled in the same
way this data has been scaled.
The sixth number (0/1) indicates whether or not the classifications
have been scaled.
The seventh number indicates the lower bound of the scaled data.
The eighth number indicates the upper bound of the scaled data.
Then follows a list of the thresholds used for scaling (double). It
has as many elements as there are dimensions in input space plus the
number of dimensions of the classification.
Then follows a list of scaling factors (double). It has as many
elements as the previous list.
For an exact explanation on how scale factors and threshold are
calculated see the section on scaling. Note that these scale factors
are the factors that have previously been applied to the data. They
will not be applied to the data when loading.
The rest of the file simply consists of examples. First the input
values of an example then its classification.

Say we have four examples in two dimensional input space and the
classification follows the function f(x_1,x_2)=2 x x_1 + x_2.
The data was scaled before being put into the list between  -1 and 1.
The original data points are the same as in the version -1 example.
The input file should look something like this:

4
2
0
1
1
0
-1
1
-0.75    -1.7       0
1.333333 0.58823529 1
0.333333 -0.4117647 3
1.5      1          6.4
1.2      -1         2.4
0        0.76470588 3


Version 1:

The fourth number (int) indicates the dimensionality of the classification of
the examples. (In this version this has to be 1.)
The fifth number (0/1) indicates whether or not the data has
individual epsilon values per example.
The sixth number (0/1) indicates whether or not the data has  been
pre-scaled. This is useful if other data shoudl be scaled in the same
way this data has been scaled.
The seventh number (0/1) indicates whether or not the classifications
have been scaled.
The eighth number indicates the lower bound of the scaled data.
The ninth number indicates the upper bound of the scaled data.
Then follows a list of the thresholds used for scaling (double). It
has as many elements as there are dimensions in input space plus the
number of dimensions of the classification.
Then follows a list of scaling factors (double). It has as many
elements as the previous list.
For an exact explanation on how scale factors and threshold are
calculated see the section on scaling. Note that these scale factors
are the factors that have previously been applied to the data. They
will not be applied to the data when loading.
The rest of the file simply consists of examples. First the input
values of an example then its classification.

Say we have four examples in two dimensional input space and the
classification follows the function f(x_1,x_2)=2 x x_1 + x_2.
The data was scaled before being put into the list between  -1 and 1.
The original data points are the same as in the version -1 example.
The input file should look something like this:

4
2
-1
1
1
1
0
0
0
0   0   0
1   1   1
1   1   3   0.1
1.5 3.4 6.4 0.2
1.2 0   2.4 0.1
0   3   3   0.2



Scaling

Scaling has to be used when values become unmanagable for the
optimizer used in the SV Machine. Some values reduce the numerical
accuracy to such an extent that no solution can be found anymore.

Scaling a set of numbers N works as follows:

We are given the lower and upper bound (lb,ub) between which the
scaling should occur.
Find the maximum and minimum value in N: max(N), min(N)
Calculate the scaling factor: s=(ub-lb)/(max(N)-min(N))
Calculate the threshold: t=(lb/s)-min(N)

Scale all samples x:

x_s=(x+t)*s


Loading and saving chunks

Chunks are very similar to sample lists except that they store the
alpha values as well as the samples.

The chunk is an ascii file only containing numbers. The first
few numbers indicate the exact format followed by the data.

The chunk can load several formats but only saves one format.

The first number (int) of the chunk file always contains the number of
examples in the file.

The second number (int) of the chunk file always contains the
dimensionality of the input space.

The third number (int) of the sample list file determines the format
of the file. If it is positive, this is the dimensionality
classification of the examples (version -1), this has to be 1 in this
version. If it is negative it indicates the version number of the file
format.

Version -1:

The fourth number (int) indicates the number of alpha values per
example.
This is followed by an example as described in the sample list
followed by the alpha values of that example and then the next example
and so on.

Version 0:

This version additionally can handle multiple epsilons.

The fourth number (int) indicates the dimensionality of the classification of
the examples. (In this version this has to be 1.)
The fifth number (int) indicates the number of alpha values per example. 
The sixth number (0/1) indicates whether or not the data has
individual epsilon values per example.

This is followed by an example as described in the sample list
followed by the alpha values of that example and then the next example
and so on.