Implemented significant signatures liblinear regression

bioclipse · Jun 16, 2015 · a9c79c8 · a9c79c8
1 parent 0bf9282
commit a9c79c8
Showing 1 changed file with 86 additions and 1 deletion.
diff --git a/....bioclipse.ds.liblinear/src/net/bioclipse/ds/liblinear/SignaturesLibLinearPrediction.java b/....bioclipse.ds.liblinear/src/net/bioclipse/ds/liblinear/SignaturesLibLinearPrediction.java
@@ -21,6 +21,7 @@
 import java.util.List;
 import java.util.Map;
 
+import libsvm.svm;
 import libsvm.svm_node;
 import net.bioclipse.cdk.domain.ICDKMolecule;
 import net.bioclipse.core.business.BioclipseException;
@@ -330,10 +331,94 @@ else if (prediction<regrLowerThreshold && !lowIsNegative)
 			match.setName("Result: " + formatter.format(prediction));
 //			return results;
 		}
+
+		Map<Integer, Double> atomGradientComponents = getAtomGradientComponents(model, instance, predModel);
+
+		match = new ScaledResultMatch("Result: " 
+				+ formatter.format( prediction ), 
+				ITestResult.INFORMATIVE);
+		results.clear();
+		results.add(match);
+
+		int result=ITestResult.INCONCLUSIVE;
+		if (regrLowerThreshold!=null && regrUpperThreshold!=null){
+
+			if (prediction<regrLowerThreshold){
+				if (lowIsNegative)
+					result=ITestResult.NEGATIVE;
+				else
+					result=ITestResult.POSITIVE;
+			}else if (prediction>regrUpperThreshold){
+				if (lowIsNegative)
+					result=ITestResult.POSITIVE;
+				else
+					result=ITestResult.NEGATIVE;
+			}
+			match.setClassification(result);
+		}
+
+
+		System.out.println("Scaled results:");
+		//Color atoms according to accumulated gradient values
+		for (int currentAtomNr : atomGradientComponents.keySet()){
+			Double currentDeriv = atomGradientComponents.get(currentAtomNr);
+
+			double scaledDeriv = scaleDerivative(currentDeriv);
+			match.putAtomResult( currentAtomNr, scaledDeriv );
+			System.out.println("Atom: " + currentAtomNr + " has deriv=" + currentDeriv +" scaled=" + scaledDeriv );
+
+		}
 
 		//No extracted signatures yet, return
 		return results;
+	}
+
+	public Map<Integer, Double> getAtomGradientComponents(Model model, Feature[] instance, PredictionModel predModel){
+
+		// Get the most significant signature for classification or the sum of all gradient components for regression.
+		List<Double> gradientComponents = new ArrayList<Double>();
+		int nOverk = fact(model.getNrClass())/(fact(2)*fact(model.getNrClass()-2)); // The number of decision functions for a classification.
+		double decValues[] = new double[nOverk];
+		double lowerPointValue[] = new double[nOverk]; 
+		double higherPointValue[] = new double[nOverk];
+		Linear.predictValues(model, instance, decValues);
+		lowerPointValue = decValues.clone();
+		for (int element = 0; element < instance.length; element++){
+			// Temporarily increase the descriptor value by one to compute the corresponding component of the gradient of the decision function.
+			instance[element].setValue(instance[element].getValue() + 1.00);
+			Linear.predictValues(model, instance, decValues);
+			higherPointValue = decValues.clone();
+			double gradComponentValue = 0.0;
+			for (int curDecisionFunc = 0; curDecisionFunc < nOverk; curDecisionFunc++) {
+				gradComponentValue = gradComponentValue + higherPointValue[curDecisionFunc]-lowerPointValue[curDecisionFunc];
+			}
+			gradientComponents.add(gradComponentValue);
+			// Set the value back to what it was.
+			instance[element].setValue(instance[element].getValue() - 1.00);
+
+		}
+
+		Map<Integer, Double> atomGreadientComponents = new HashMap<Integer, Double>(); // Contains a sum of all gradient components, based on modelSignatures, for a given atom.
+		for (int element = 0; element < instance.length; element++){
+			double componentVal = gradientComponents.get(element);
+			List<Integer> atomNrList = predModel.getMoleculeSignaturesAtomNr().get(signLibLinearModel.getModelSignatures().get(instance[element].getIndex()-1));
+			Iterator<Integer> atomNrInteger = atomNrList.iterator();
+			while (atomNrInteger.hasNext()){
+				int atomNr = atomNrInteger.next();
+				if (atomGreadientComponents.containsKey(atomNr)){
+					atomGreadientComponents.put(atomNr, atomGreadientComponents.get(atomNr)+componentVal);
+				}
+				else{
+					atomGreadientComponents.put(atomNr,componentVal);	
+				}
+			}
+		}
+		System.out.println(atomGreadientComponents.toString());					
 
+
+		return atomGreadientComponents;
+	}
+
 
 //		//End here if we should not extract significant signatures
 //		if (!extractSignificantSignatures)
@@ -566,7 +651,7 @@ else if (prediction<regrLowerThreshold && !lowIsNegative)
 //		}
 //
 //		return results;
-	}
+//	}
 
 	/**
 	 * A prediction model tales a mol as input, sets up all signatures needed,