svn commit: r1160069 - /mahout/trunk/core/src/main/java/org/apache/mahout/classifier/sgd/MixedGradient.java

[sr...@apache.org]

Author: srowen
Date: Sun Aug 21 21:44:14 2011
New Revision: 1160069

URL: http://svn.apache.org/viewvc?rev=1160069&view=rev
Log:
MAHOUT-678 commit reasonable band-aid proposed in the patch to resolve the proximate issue: no more NPE when not running on binomial data

Modified:
    mahout/trunk/core/src/main/java/org/apache/mahout/classifier/sgd/MixedGradient.java

Modified: mahout/trunk/core/src/main/java/org/apache/mahout/classifier/sgd/MixedGradient.java
URL: http://svn.apache.org/viewvc/mahout/trunk/core/src/main/java/org/apache/mahout/classifier/sgd/MixedGradient.java?rev=1160069&r1=1160068&r2=1160069&view=diff
==============================================================================
--- mahout/trunk/core/src/main/java/org/apache/mahout/classifier/sgd/MixedGradient.java (original)
+++ mahout/trunk/core/src/main/java/org/apache/mahout/classifier/sgd/MixedGradient.java Sun Aug 21 21:44:14 2011
@@ -24,11 +24,13 @@ import org.apache.mahout.math.Vector;
 import java.util.Random;
 
 /**
- * Provides a stochastic mixture of ranking updates and normal logistic updates. This uses a
+ * <p>Provides a stochastic mixture of ranking updates and normal logistic updates. This uses a
  * combination of AUC driven learning to improve ranking performance and traditional log-loss driven
- * learning to improve log-likelihood.
- * <p/>
- * See www.eecs.tufts.edu/~dsculley/papers/combined-ranking-and-regression.pdf
+ * learning to improve log-likelihood.</p>
+ *
+ * <p>See www.eecs.tufts.edu/~dsculley/papers/combined-ranking-and-regression.pdf</p>
+ *
+ * <p>This implementation only makes sense for the binomial case.</p>
  */
 public class MixedGradient implements Gradient {
 
@@ -36,6 +38,8 @@ public class MixedGradient implements Gr
   private final RankingGradient rank;
   private final Gradient basic;
   private final Random random = RandomUtils.getRandom();
+  private boolean hasZero;
+  private boolean hasOne;
 
   public MixedGradient(double alpha, int window) {
     this.alpha = alpha;
@@ -47,8 +51,13 @@ public class MixedGradient implements Gr
   public Vector apply(String groupKey, int actual, Vector instance, AbstractVectorClassifier classifier) {
     if (random.nextDouble() < alpha) {
       // one option is to apply a ranking update relative to our recent history
+      if (!hasZero || !hasOne) {
+        throw new IllegalStateException();
+      }
       return rank.apply(groupKey, actual, instance, classifier);
     } else {
+      hasZero |= actual == 0;
+      hasOne |= actual == 1;
       // the other option is a normal update, but we have to update our history on the way
       rank.addToHistory(actual, instance);
       return basic.apply(groupKey, actual, instance, classifier);