"writing an Input Filter for One-Class SVM"

niccay

Hello,

I'm using a dataset with a binary (nominal) label property I'd like to classify by a One-Class SVM. In my project, I'm importing my dataset with a csv Import Operator and using this as Input for a X-validation operator.

X-Validation contains just a simple One-Class (Lib)SVM Learner and for testing a Apply Model+Performance Operator.

Unfortunately the One ClassLibSVM Learner expects my trainingsset to just have 1 label value (I'm having 2 different values..), thus I thought writing a Filter throwing away positive(minority) class examples should be sufficient for this task (see source below). Unfortunately it doesn't. Well, I think I've to fix the Nominal Mapping of my label attribute but I'm not sure where to start and what to do.

Does anyone has a clue how to solve my issue? (But please do me a little favor and don't refer to the whitepaper)

Thanks,
Nico

	 /*
	 * 
	 * Removes minority class examples of exampleset in order to use an one-class SVM
	 * 
	 * */
	public ExampleSet apply(ExampleSet trainingsset) throws OperatorException {
		List<Example> majorityExamples = new LinkedList<Example>();
		NominalAttribute label = (NominalAttribute) trainingsset.getAttributes().getLabel();
		double minorityLabelValue = getMinorityClass(trainingsset);


		// only care iff exampleset contains 2 different classes
		if(label.isNominal() && label.getMapping().size()==2) 

			for (Example e : trainingsset)
			{
				if(e.getLabel()!=minorityLabelValue)
					majorityExamples.add(e);

			}

		// NOT WORKING :-(
		//NominalMapping labelMapping = label.getMapping();
		//(String majorityLabelMapping = label.getMapping().getNegativeString();
		//labelMapping.clear();
		//labelMapping.setMapping(majorityLabelMapping, 0);

		return computeExampleSet(majorityExamples);
	}

protected ExampleSet computeExampleSet(List<Example> examples) 
	{
		int anzBsp = examples.size();
		int anzAtr = examples.get(0).getAttributes().size();
		double[][] values = new double[anzBsp][anzAtr];
		double[] labels = new double[anzBsp];
		String[] attrNames = new String[anzAtr];
		String labelName;
		int i=0,n=0;

		for(Attribute a : examples.get(0).getAttributes()) {
			attrNames = a.getName();
			i++;
		}
		Attribute label = examples.get(0).getAttributes().getLabel();
		labelName = label.getName();
		i=0;
		for(Example e : examples) {
			n=0;
			for(Attribute a : e.getAttributes()) {
				values = e.getValue(a);
				n++;
			}
			labels = e.getLabel();
			i++;
		}

		return ExtendedExampleSetFactory.createExampleSet(values,labels,attrNames,labelName,label.getMapping());


	}

public static ExampleSet createExampleSet(double[][] data, double[] labels, String[] attrNames, String labelName, NominalMapping lblMapping) {
	        if (data.length == 0) {
	            throw new RuntimeException("ExampleSetFactory.createExampleSet(double[][], double[]): data matrix is not allowed to be empty.");
	        }
	                
	        // create attributes
	        int numberOfAttributes = data[0].length;
	        List<Attribute> attributeList = new ArrayList<Attribute>(numberOfAttributes + (labels != null ? 1 : 0));
	        for (int a = 0; a < numberOfAttributes; a++) {
	            attributeList.add(AttributeFactory.createAttribute(attrNames, Ontology.NUMERICAL));
	        }
	        Attribute labelAttribute = null;
	        
	        if (labels != null) {
	            labelAttribute = AttributeFactory.createAttribute(labelName, Ontology.NOMINAL);
	            labelAttribute.setMapping(lblMapping);
	            attributeList.add(labelAttribute);
	        }

	        
	        // create table
	        MemoryExampleTable table = new MemoryExampleTable(attributeList);
	        for (int e = 0; e < data.length; e++) {
	            double[] dataRow = data;
	            if (labelAttribute != null) {
	                dataRow = new double[numberOfAttributes + 1];
	                System.arraycopy(data, 0, dataRow, 0, data.length);
	                dataRow[dataRow.length - 1] = labels;
	            }
	            table.addDataRow(new DoubleArrayDataRow(dataRow));
	        }
	        
	        return table.createExampleSet(labelAttribute);
	    }

land

Hi,
actually there's a whitepaper... :P Just a joke.

Check out the CommunityExtension and download the process about one class svms. Marco Stolpe describe in another thread, that he has created a process for training and validating a one class SVM on a trainingset with binominal label.

Greetings,
  Sebastian