Improve robustness of PAM initialization. Fixes Github#4

elki-project · Aug 3, 2015 · b48eb01 · b48eb01
1 parent a5bb9f1
commit b48eb01
Show file tree

Hide file tree

Showing 2 changed files with 22 additions and 20 deletions.
diff --git a/elki/src/main/java/de/lmu/ifi/dbs/elki/algorithm/clustering/kmeans/KMedoidsPAM.java b/elki/src/main/java/de/lmu/ifi/dbs/elki/algorithm/clustering/kmeans/KMedoidsPAM.java
@@ -46,19 +46,19 @@
 /**
  * The original PAM algorithm or k-medoids clustering, as proposed by Kaufman
  * and Rousseeuw in "Partitioning Around Medoids".
- * 
+ *
  * Reference:
  * <p>
  * Clustering my means of Medoids<br />
  * Kaufman, L. and Rousseeuw, P.J.<br />
  * in: Statistical Data Analysis Based on the L1-Norm and Related Methods
  * </p>
- * 
+ *
  * @author Erich Schubert
- * 
+ *
  * @apiviz.has MedoidModel
  * @apiviz.composedOf KMedoidsInitialization
- * 
+ *
  * @param <V> vector datatype
  */
 @Title("Partioning Around Medoids")
@@ -93,7 +93,7 @@ public class KMedoidsPAM<V> extends AbstractDistanceBasedAlgorithm<V, Clustering
 
   /**
    * Constructor.
-   * 
+   *
    * @param distanceFunction distance function
    * @param k k parameter
    * @param maxiter Maxiter parameter
@@ -108,7 +108,7 @@ public KMedoidsPAM(DistanceFunction<? super V> distanceFunction, int k, int maxi
 
   /**
    * Run k-medoids
-   * 
+   *
    * @param database Database
    * @param relation relation to use
    * @return result
@@ -155,7 +155,7 @@ public Clustering<MedoidModel> run(Database database, Relation<V> relation) {
 
   /**
    * Run the PAM optimization phase.
-   * 
+   *
    * @param distQ Distance query
    * @param ids IDs to process
    * @param medoids Medoids list
@@ -268,7 +268,7 @@ protected void runPAMOptimization(DistanceQuery<V> distQ, DBIDs ids, ArrayModifi
   /**
    * Returns a list of clusters. The k<sup>th</sup> cluster contains the ids of
    * those FeatureVectors, that are nearest to the k<sup>th</sup> mean.
-   * 
+   *
    * @param means Object centroids
    * @param ids Object ids
    * @param nearest Distance to nearest medoid
@@ -296,6 +296,9 @@ else if(dist < mindist2) {
           mindist2 = dist;
         }
       }
+      if(minIndex < 0) {
+        throw new AbortException("Too many infinite distances. Cannot assign objects.");
+      }
       assignment.put(iditer, minIndex);
       nearest.put(iditer, mindist);
       second.put(iditer, mindist2);
@@ -316,9 +319,9 @@ protected Logging getLogger() {
 
   /**
    * Parameterization class.
-   * 
+   *
    * @author Erich Schubert
-   * 
+   *
    * @apiviz.exclude
    */
   public static class Parameterizer<V> extends AbstractDistanceBasedAlgorithm.Parameterizer<V> {

diff --git a/.../java/de/lmu/ifi/dbs/elki/algorithm/clustering/kmeans/initialization/PAMInitialMeans.java b/.../java/de/lmu/ifi/dbs/elki/algorithm/clustering/kmeans/initialization/PAMInitialMeans.java
@@ -47,18 +47,16 @@
 
 /**
  * PAM initialization for k-means (and of course, PAM).
- * 
+ *
  * Reference:
  * <p>
  * Clustering my means of Medoids<br />
  * Kaufman, L. and Rousseeuw, P.J.<br />
  * in: Statistical Data Analysis Based on the L_1–Norm and Related Methods
  * </p>
- * 
- * TODO: enforce using a distance matrix?
- * 
+ *
  * @author Erich Schubert
- * 
+ *
  * @param <O> Object type for KMedoids initialization
  */
 @Reference(title = "Clustering my means of Medoids", //
@@ -98,11 +96,11 @@ public <T extends NumberVector, V extends NumberVector> List<V> chooseInitialMea
   public DBIDs chooseInitialMedoids(int k, DBIDs ids, DistanceQuery<? super O> distQ) {
     ArrayModifiableDBIDs medids = DBIDUtil.newArray(k);
     DBIDVar bestid = DBIDUtil.newVar();
-    double best = Double.POSITIVE_INFINITY;
     WritableDoubleDataStore mindist = null;
 
     // First mean is chosen by having the smallest distance sum to all others.
     {
+      double best = Double.POSITIVE_INFINITY;
       WritableDoubleDataStore newd = null;
       FiniteProgress prog = LOG.isVerbose() ? new FiniteProgress("Choosing initial mean", ids.size(), LOG) : null;
       for(DBIDIter iter = ids.iter(); iter.valid(); iter.advance()) {
@@ -138,6 +136,7 @@ public DBIDs chooseInitialMedoids(int k, DBIDs ids, DistanceQuery<? super O> dis
     FiniteProgress prog = LOG.isVerbose() ? new FiniteProgress("Choosing initial centers", k, LOG) : null;
     LOG.incrementProcessed(prog); // First one was just chosen.
     for(int i = 1; i < k; i++) {
+      double best = Double.POSITIVE_INFINITY;
       WritableDoubleDataStore bestd = null, newd = null;
       for(DBIDIter iter = ids.iter(); iter.valid(); iter.advance()) {
         if(medids.contains(iter)) {
@@ -162,8 +161,8 @@ public DBIDs chooseInitialMedoids(int k, DBIDs ids, DistanceQuery<? super O> dis
           newd = null;
         }
       }
-      if(bestid == null) {
-        throw new AbortException("No median found that improves the criterion function?!?");
+      if(bestd == null) {
+        throw new AbortException("No median found that improves the criterion function?!? Too many infinite distances.");
       }
       medids.add(bestid);
       if(newd != null) {
@@ -181,9 +180,9 @@ public DBIDs chooseInitialMedoids(int k, DBIDs ids, DistanceQuery<? super O> dis
 
   /**
    * Parameterization class.
-   * 
+   *
    * @author Erich Schubert
-   * 
+   *
    * @apiviz.exclude
    */
   public static class Parameterizer<V> extends AbstractParameterizer {