Changeset 140 for trunk/CrossPare/src/de/ugoe/cs/cpdp/training/MetricMatchingTraining.java

Timestamp:

08/22/16 12:02:26 (8 years ago)

Author:

atrautsch

Message:

More cleanup and comments

File:

• trunk/CrossPare/src/de/ugoe/cs/cpdp/training/MetricMatchingTraining.java (modified) (9 diffs)

trunk/CrossPare/src/de/ugoe/cs/cpdp/training/MetricMatchingTraining.java

@@ -49 +49 @@
  * This also means we can use any Weka Classifier not just LogisticRegression.
  * 
+ * Config: <setwisetestdataawaretrainer name="MetricMatchingTraining" param="Logistic weka.classifiers.functions.Logistic" threshold="0.05" method="spearman"/>
+ * Instead of spearman metchod it also takes ks, percentile.
+ * Instead of Logistic every other weka classifier can be chosen.
+ * 
+ * Future work:
+ * implement chisquare test in addition to significance for attribute selection
+ * http://commons.apache.org/proper/commons-math/apidocs/org/apache/commons/math3/stat/inference/ChiSquareTest.html
+ * use chiSquareTestDataSetsComparison
  */
 public class MetricMatchingTraining extends WekaBaseTraining implements ISetWiseTestdataAwareTrainingStrategy {
 
     private MetricMatch mm = null;
-    private Classifier classifier = new MetricMatchingClassifier();
+    private Classifier classifier = null;
     
     private String method;
@@ -89 +97 @@
         @Override
         public void apply(SetUniqueList<Instances> traindataSet, Instances testdata) {
-
+            // reset these for each run
+            this.mm = null;
+            this.classifier = null;
+            
                 double score = 0; // matching score to select the best matching training data from the set
                 int num = 0;
@@ -116 +127 @@
                 }
                 
-                // if we have found a matching instance we use it
+                // if we have found a matching instance we use it, log information about the match for additional eval later
                 Instances ilist = null;
                 if (this.mm != null) {
@@ -131 +142 @@
                 try {
                         if(this.mm != null) {
+                            this.classifier = new MetricMatchingClassifier();
                             this.classifier.buildClassifier(ilist);
                             ((MetricMatchingClassifier) this.classifier).setMetricMatching(this.mm);
@@ -257 +269 @@
                     return this.train_values.get(0).length;
                 }
-                
-                /** 
-                 * This creates a new Instance out of the passed Instance and the previously matched attributes.
-                 * We do this because the evaluation phase requires an original Instance with every attribute.
+
+                
+                /**
+                 * The test instance must be of the same dataset as the train data, otherwise WekaEvaluation will die.
+                 * This means we have to force the dataset of this.train (after matching) and only 
+                 * set the values for the attributes we matched but with the index of the traindata attributes we matched.
                  * 
-                 * @param test instance
-                 * @return new instance
+                 * @param test
+                 * @return
                  */
                 public Instance getMatchedTestInstance(Instance test) {
-                    //create new instance with our matched number of attributes + 1 (the class attribute)
-                        Instances testdata = this.getMatchedTest();
-
-                        Instance ni = new DenseInstance(this.attributes.size()+1);
-                        ni.setDataset(testdata);
-                         
-                for(Map.Entry<Integer, Integer> attmatch : this.attributes.entrySet()) {
-                    ni.setValue(testdata.attribute(attmatch.getKey()), test.value(attmatch.getValue()));
-                }
-                
-                        ni.setClassValue(test.value(test.classAttribute()));
-
-                        return ni;
-        }
-
+            Instance ni = new DenseInstance(this.attributes.size()+1);
+            
+            Instances inst = this.getMatchedTrain();
+            
+            ni.setDataset(inst);
+            
+            // assign only the matched attributes to new indexes
+            double val;
+            int k = 0;
+            for(Map.Entry<Integer, Integer> attmatch : this.attributes.entrySet()) {
+                // get value from matched attribute
+                val = test.value(attmatch.getValue());
+                
+                // set it to new index, the order of the attributes is the same
+                ni.setValue(k, val);
+                k++;
+            }
+            ni.setClassValue(test.value(test.classAttribute()));
+
+            return ni;
+                }
+
+                
         /**
          * returns a new instances array with the metric matched training data
@@ -494 +516 @@
                 // -1 means that it is not in the set of maximal matching
                 if( i != -1 && result[i] != -1) {
-                    //Console.traceln(Level.INFO, "Found maximal bipartite match between: "+ i + " and " + result[i]);
+                    this.p_sum += mwbm.weights[i][result[i]];  // we add the weight of the returned matching for scoring the complete match later
                     this.attributes.put(i, result[i]);
                 }
@@ -544 +566 @@
                     
                     if( score > cutoff ) {
-                        this.p_sum += score;
                         mwbm.setWeight(i, j, score);
                     }
@@ -587 +608 @@
                                         p = t.correlation(this.train_values.get(i), this.test_values.get(j));
                                         if (p > cutoff) {
-                                            this.p_sum += p;
                         mwbm.setWeight(i, j, p);
                                         }
@@ -667 +687 @@
                                         p = t.approximateP(t.kolmogorovSmirnovStatistic(this.train_values.get(i), this.test_values.get(j)), this.train_values.get(i).length, this.test_values.get(j).length);
                                         if (p > cutoff) {
-                        this.p_sum += p;
                         mwbm.setWeight(i, j, p);
                                         }