reduce_class_svmreduce_class_svmReduceClassSvmReduceClassSvm (Operator)

Name

reduce_class_svmreduce_class_svmReduceClassSvmReduceClassSvm — Approximate a trained support vector machine by a reduced support vector machine for faster classification.

Signature

reduce_class_svm( : : SVMHandle, Method, MinRemainingSV, MaxError : SVMHandleReduced)

As described in create_class_svmcreate_class_svmCreateClassSvmCreateClassSvmCreateClassSvm, the classification time of a SVM depends on the number of kernel evaluations between the support vectors and the feature vectors. While the length of the data vectors can be reduced in a preprocessing step like 'principal_components'"principal_components""principal_components""principal_components""principal_components" or 'canonical_variates'"canonical_variates""canonical_variates""canonical_variates""canonical_variates" (see create_class_svmcreate_class_svmCreateClassSvmCreateClassSvmCreateClassSvm for details), the number of resulting SV depends on the complexity of the classification problem. The number of SVs is determined during training. To further reduce classification time, the number of SVs can be reduced by approximating the original separating hyperplane with fewer SVs than originally required. For this purpose, a copy of the original SVM provided by SVMHandleSVMHandleSVMHandleSVMHandleSVMHandle is created and returned in SVMHandleReducedSVMHandleReducedSVMHandleReducedSVMHandleReducedSVMHandleReduced. This new SVM has the same parametrization as the original SVM, but a different SV expansion. The training samples that are included in SVMHandleSVMHandleSVMHandleSVMHandleSVMHandle are not copied. The original SVM is not modified by reduce_class_svmreduce_class_svmReduceClassSvmReduceClassSvmReduceClassSvm.

The reduction method is selected with MethodMethodMethodMethodmethod. Currently, only a bottom up approach is supported, which iteratively merges SVs. The algorithm stops if either the minimum number of SVs is reached (MinRemainingSVMinRemainingSVMinRemainingSVMinRemainingSVminRemainingSV) or if the accumulated maximum error exceeds the threshold MaxErrorMaxErrorMaxErrorMaxErrormaxError. Note that the approximation reduces the complexity of the hyperplane and thereby leads to a deteriorated classification rate. A common approch is therefore to start from a small MaxErrorMaxErrorMaxErrorMaxErrormaxError e.g., 0.001, and to increase its value step by step. To control the reduction ratio, at each step the number of remaining SVs is determined with get_support_vector_num_class_svmget_support_vector_num_class_svmGetSupportVectorNumClassSvmGetSupportVectorNumClassSvmGetSupportVectorNumClassSvm and the classification rate is checked on a separate test data set with classify_class_svmclassify_class_svmClassifyClassSvmClassifyClassSvmClassifyClassSvm.

Execution Information

Multithreading type: reentrant (runs in parallel with non-exclusive operators).
Multithreading scope: global (may be called from any thread).
Processed without parallelization.

Parameters

SVMHandleSVMHandleSVMHandleSVMHandleSVMHandle (input_control) class_svm → (integer)

Original SVM handle.

MethodMethodMethodMethodmethod (input_control) string → (string)

Type of postprocessing to reduce number of SV.

Default value: 'bottom_up' "bottom_up" "bottom_up" "bottom_up" "bottom_up"

List of values: 'bottom_up'"bottom_up""bottom_up""bottom_up""bottom_up"

MinRemainingSVMinRemainingSVMinRemainingSVMinRemainingSVminRemainingSV (input_control) integer → (integer)

Minimum number of remaining SVs.

Default value: 2

Suggested values: 2, 3, 4, 5, 7, 10, 15, 20, 30, 50

Restriction: MinRemainingSV >= 2

MaxErrorMaxErrorMaxErrorMaxErrormaxError (input_control) real → (real)

Maximum allowed error of reduction.

Default value: 0.001

Suggested values: 0.0001, 0.0002, 0.0005, 0.001, 0.002, 0.005, 0.01, 0.02, 0.05

Restriction: MaxError > 0.0

SVMHandleReducedSVMHandleReducedSVMHandleReducedSVMHandleReducedSVMHandleReduced (output_control) class_svm → (integer)

SVMHandle of reduced SVM.

Example (HDevelop)

* Train an SVM
create_class_svm (NumFeatures, 'rbf', 0.01, 0.01, NumClasses,\
                  'one-versus-all', 'normalization', NumFeatures,\
                  SVMHandle)
read_samples_class_svm (SVMHandle, 'samples.mtf')
train_class_svm (SVMHandle, 0.001, 'default')
* Create a reduced SVM
reduce_class_svm (SVMHandle, 'bottom_up', 2, 0.01, SVMHandleReduced)
write_class_svm (SVMHandleReduced, 'classifier.svm')
clear_class_svm (SVMHandleReduced)
clear_class_svm (SVMHandle)

Result

If the parameters are valid the operator train_class_svmtrain_class_svmTrainClassSvmTrainClassSvmTrainClassSvm returns the value 2 (H_MSG_TRUE). If necessary, an exception is raised.