Hyper Classifier Example

Warning
Preparing the data
Initializing and training the classifier
Getting a response
Further information

Warning

Expect this file to take a relatively long time to execute due to expensive calculations. Please be patient.

Preparing the data

This example comes with two files, 'simpleCurveSegments.mat' and 'realCurveShapeAnalystSamples400_.mat', which contain basic curve samples for the lowest level classifier and real curve samples respectively. All samples are associated with a correct classification id.

load('simpleCurveSegments.mat');
load('realCurveShapeAnalystSamples400_.mat');

Initializing and training the classifier

As the algorithm requires a simple LCC classifier at its lowest level, we create one and train it using the basic curve samples.

nDCT = 0;
nCV  = [ 3 5 7 9 11 15 17 ];

config.scalePercentage= 0.3;
config.growthThresh= 0.95;
config.keepHistory= 0;

config = createClassifierConfig(config);
classifier = newClassifier(protos3CV, protosIds, [], config);

This LCC classifier is then used to build the hyper classifier.

configH.filterClassifier = classifier;

We reset the previous basic configuration object and recreate it to have a clean start with specific parameters. Within the hyper classifier this new configuration will be used as a template for the higher level LCC classifiers.

clear config;

config.scalePercentage    = 0.3;
config.growthThresh       = 0.95;
config.keepHistory        = 0;
config.generateNewClasses = 1;
config.recycleRubbish     = 0;
config.scaleAdaptation    = 'allMean';

configH.classifierConfigs = createClassifierConfig(config);

Finally the actual hyper classifier can be created, following the usual steps:

1. Specify non-default parameters in a structure

configH.nHierarchyLevels = maxLevel;

for lev = 1 : maxLevel
    configH.featureParams(lev).nFeatures = [ 0 0 nCV(lev) ];
end

2. Create complete configuration object for hyper classifiers

configH = createHyperClassifierConfig(configH);

3. Create hyper classifier

hyperClassif = newHyperClassifier(samples, classIds, configH);

..........
..........
..........
..........

Getting a response

Although the algorithm is capable of considering multiple hypotheses and thus achieve better results, we will use only one hypothesis here. However, in general the responses are returned in a matrix with one row for each considered hypothesis.

nHypos = 1;

[ responses, responseValues ] = getHyperClassifierResponse(hyperClassif, samples, nHypos);

The quality of the response can easily be evaluated, since the data we used here has already been pre-classified.

nFalse = sum((classIds - responses(1, :)) ~= 0);

fprintf('false classified: %d\n', nFalse);
fprintf('false classified: %5.3f\n\n', nFalse / length(samples));

false classified: 0
false classified: 0.000

Further information

Details about the LCC classifiers which are used to construct the hyper classifier can be found in the LCC standalone package.