SVMTrain

Function | Description | Signature | Input Arguments | Output Arguments | See Also

Function:

SVMTrain

Description:

Construct a SVM classifier or regressor.
- This function performs input parameter checking and if successful, calls mexSVMTrain for training.

Signature:

[AlphaY, SVs, Bias, Parameters, nSV, nLabel] = SVMTrain(Samples, Labels, Parameters, Weight)

Input Arguments:

Name:	Optional:	Description:
Samples	No	The N input patterns in M dimensions (MxN matrix, a row vector of columns).
Labels	No	The labels of the N input patterns (1xN row vector).
Parameters	Yes	The parameters to control training (1xa row vector, a<=11). This row vector must consist of the following elements. (1) Kernel Type: (default=2). (0) Linear (1) Polynomial: Gamma(<X(:,i),X(:,j)>+Coefficient)^Degree (2) RBF: (exp(-Gamma\|X(:,i)-X(:,j)\|^2)) (3) Sigmoid: tanh(Gamma*<X(:,i),X(:,j)>+Coefficient) (2) Degree: (default=3). (3) Gamma: If the input value is zero, Gamma will be set to 1.0 /(max_pattern_dimension). Otherwise, Gamma will remain unchanged in the function (default=1). (4) Coefficient: (default=0). (5) C: Cost of constraint violation for C-SVC, epsilon-SVR, and nu-SVR (default=1). (6) Cache: Space to hold the elements of K(<X(:,i),X(:,j)>) matrix (default=40MB). (7) epsilon: tolerance of termination criterion (default=0.001). (8) SVM Type (default=0) (0) C-SVC (1) nu-SVC (2) one-class SVM (3) epsilon-SVR (4) nu-SVR (9) nu: nu of nu-SVC, one-class SVM, and nu-SVR (default=0.5). (10) loss tolerance: epsilon in loss function of epsilon-SVR (default=0.1). (11) shrinking: use shrinking heuristics. (default=1<yes>).
Weight	Yes	Used to weight classes. This is acheived by multiplying the C for a class i by weight(i) in C-SVC (default=all 1.0 's).

Output Arguments:

Name:	Optional:	Description:
AlphaY	No	Alpha * Y where the Alpha vector contains the non-zero Lagrange Coefficients and Y are the corresponding labels. Binary Classification/Outlier Detection: A row vector where AlphaY(i) corresponds to the multiplier of the i'th support vector. Multi-Class Classification: A (L-1) x sum(nSV) matrix. AlphaY(i,j) is the multiplier for a classifier between classes i and (i+1) where i < length(nSV) and corresponds to the j'th support vector. Regression: A row vector where AlphaY(i) corresponds to the i'th support vector.
SVs	No	The Support Vectors (M x sum(nSV) matrix) for this classifier/regressor-- the samples corresponding to the non-zero Lagrange multipliers AlphaY. A M x sum(nSV) matrix of sum(nSV) support vectors in M dimensions. Binary Classification/Outlier Detection/Multi-class Classification: The class of a support vector i is obtained by: nLabel(min(find(cumsum(nSV)>i))). The columns from u(i) to u(i+1) represent the support vectors with a corresponding class nLabel(i) where u=[1,cumsum(nSV)]. [SVs from Class_1, SVs from Class_2, [...], SVs from Class_L]; Regression: A M x sum(nSV) matrix of support vectors. Labels do not necessarily dictate the organization of this matrix.
Bias	No	Binary Classification/Outlier Detection/Regression: A scalar value representing the bias or threshold of the SVM classifier. Multi-class Classification: A (1x(L*(L-1)/2)) row vector of biases for all the two-class classifiers.
Parameters	Yes	The training parameters used.
nSV	No	The numbers of SVs for each class (1xL row vector).
nLabel	Yes	The labels of each class (1xL row vector).

See Also:

Portions Copyright (C) by Contributors to the OSU SVM Package, 2000-2003 under the BSD License.

Generated by matdoc written by Damian Eads.