select_feature_set_svm — Selects an optimal combination of features to classify the provided data.
select_feature_set_svm selects an optimal subset from a set of features to solve a given classification problem. The classification problem has to be specified with annotated training data in ClassTrainDataHandle and will be classified by a support vector machine (SVM). Details of the properties of this classifier can be found in create_class_svm.
The result of the operator is a trained classifier that is returned in SVMHandle. Additionally, the list of indices or names of the selected features is returned in SelectedFeatureIndices. To use this classifier, calculate for new input data all features mentioned in SelectedFeatureIndices and pass them to the classifier.
A possible application of this operator can be a comparison of different parameter sets for certain feature extraction techniques. Another application is to search for a feature that is discriminating between different classes.
Additionally, the values for 'nu' and 'gamma' can be estimated for the SVM. To only estimate these two parameters without altering the feature set, the feature vector has to be specified as one large subfeature.
To define the features that should be selected from ClassTrainDataHandle, the dimensions of the feature vectors in ClassTrainDataHandle can be grouped into subfeatures by calling set_feature_lengths_class_train_data. A subfeature can contain several subsequent elements of a feature vector. The operator decides for each of these subfeatures, if it is better to use it for the classification or leave it out.
The indices of the selected subfeatures are returned in SelectedFeatureIndices. If names were set in set_feature_lengths_class_train_data, these names are returned instead of the indices. If set_feature_lengths_class_train_data was not called for ClassTrainDataHandle before, each element of the feature vector is considered as a subfeature.
The selection method SelectionMethod is either a greedy search 'greedy' (iteratively add the feature with highest gain) or the dynamically oscillating search 'greedy_oscillating' (add the feature with highest gain and test then if any of the already added features can be left out without great loss). The method 'greedy' is generally preferable, since it is faster. Only in cases when the subfeatures are low-dimensional or redundant, the method 'greedy_oscillating' should be chosen.
The optimization criterion is the classification rate of a two-fold cross-validation of the training data. The best achieved value is returned in Score.
The parameters 'nu' and 'gamma' for the SVM that is used to classify can be set to 'auto' by using the parameters GenParamNames and GenParamValues. If they are set to 'auto', the estimated optimal 'nu' and/or 'gamma' is estimated. The automatic estimation of 'nu' and 'gamma' can take a substantial amount of time (up to days, depending on the data set and the number of features).
Additionally, there is the parameter 'mode' which can be either set to 'one-versus-all' or 'one-versus-one'. An explanation of the two modes as well as of the parameters 'nu' and 'gamma' as the kernel parameter of the radial basis function (RBF) kernel can be found in create_class_svm.
This operator may take considerable time, depending on the size of the data set in the training file, and the number of features.
Please note, that this operator should not be called, if only a small set of training data is available. Due to the risk of overfitting the operator select_feature_set_svm may deliver a classifier with a very high score. However, the classifier may perfom poorly when tested.
This operator returns a handle. Note that the state of an instance of this handle type may be changed by specific operators even though the handle is used as an input parameter by those operators.
Handle of the training data.
Method to perform the selection.
Default value: 'greedy'
List of values: 'greedy', 'greedy_oscillating'
Names of generic parameters to configure the selection process and the classifier.
Default value: 
List of values: 'gamma', 'mode', 'nu'
Values of generic parameters to configure the selection process and the classifier.
Default value: 
Suggested values: 0.02, 0.05, 'auto', 'one-versus-one', 'one-versus-all'
A trained SVM classifier using only the selected features.
The selected feature set, contains indices.
The achieved score using two-fold cross-validation.
* Find out which of the two features distinguishes two Classes NameFeature1 := 'Good Feature' NameFeature2 := 'Bad Feature' LengthFeature1 := 3 LengthFeature2 := 2 * Create training data create_class_train_data (LengthFeature1+LengthFeature2,\ ClassTrainDataHandle) * Define the features which are in the training data set_feature_lengths_class_train_data (ClassTrainDataHandle, [LengthFeature1,\ LengthFeature2], [NameFeature1, NameFeature2]) * Add training data * |Feat1| |Feat2| add_sample_class_train_data (ClassTrainDataHandle, 'row', [1,1,1, 2,1 ], 0) add_sample_class_train_data (ClassTrainDataHandle, 'row', [2,2,2, 2,1 ], 1) add_sample_class_train_data (ClassTrainDataHandle, 'row', [1,1,1, 3,4 ], 0) add_sample_class_train_data (ClassTrainDataHandle, 'row', [2,2,2, 3,4 ], 1) * Add more data * ... * Select the better feature with a SVM select_feature_set_svm (ClassTrainDataHandle, 'greedy', , , SVMHandle,\ SelectedFeatureSVM, Score) clear_class_train_data (ClassTrainDataHandle) * Use the classifier * ... clear_class_svm (SVMHandle)
If the parameters are valid, the operator select_feature_set_svm returns the value 2 (H_MSG_TRUE). If necessary, an exception is raised.
create_class_train_data, add_sample_class_train_data, set_feature_lengths_class_train_data
select_feature_set_mlp, select_feature_set_knn, select_feature_set_gmm
select_feature_set_trainf_svm, gray_features, region_features