77import water .api .DocGen ;
88import water .api .Progress2 ;
99import water .api .Request ;
10- import water .api .Request .API ;
11- import water .api .Request .Default ;
1210import water .fvec .Chunk ;
1311import water .fvec .Frame ;
1412import water .fvec .NewChunk ;
@@ -60,12 +58,13 @@ public KMeans2() {
6058 if ( sourceArg != null )
6159 sourceKey = Key .make (sourceArg );
6260
63- // Drop ignored cols and, if users asks for it, cols with too many NAs
61+ // Drop ignored cols and, if user asks for it, cols with too many NAs
6462 Frame fr = DataInfo .prepareFrame (source , ignored_cols , false , false , drop_na_cols );
6563 String [] names = fr .names ();
6664 Vec [] vecs = fr .vecs ();
6765 if (vecs == null || vecs .length == 0 )
6866 throw new IllegalArgumentException ("No columns selected. Check that selected columns have not been dropped due to too many NAs." );
67+ DataInfo dinfo = new DataInfo (fr , 0 , false , normalize , false );
6968
7069 // Fill-in response based on K99
7170 String [] domain = new String [k ];
@@ -84,7 +83,7 @@ public KMeans2() {
8483 means [i ] = (float ) vecs [i ].mean ();
8584 if ( mults != null ) {
8685 double sigma = vecs [i ].sigma ();
87- mults [i ] = normalize (sigma ) ? 1 / sigma : 1 ;
86+ mults [i ] = normalize (sigma ) ? 1.0 / sigma : 1.0 ;
8887 }
8988 }
9089
@@ -94,7 +93,8 @@ public KMeans2() {
9493 if ( initialization == Initialization .None ) {
9594 // Initialize all clusters to random rows
9695 clusters = new double [k ][vecs .length ];
97- for (double [] cluster : clusters ) randomRow (vecs , rand , cluster , means , mults );
96+ for (double [] cluster : clusters )
97+ randomRow (vecs , rand , cluster , means , mults );
9898 } else {
9999 // Initialize first cluster to random row
100100 clusters = new double [1 ][];
@@ -136,7 +136,10 @@ public KMeans2() {
136136 task ._clusters = clusters ;
137137 task ._means = means ;
138138 task ._mults = mults ;
139+ task ._ncats = dinfo ._cats ;
140+ task ._nnums = dinfo ._nums ;
139141 task .doAll (vecs );
142+
140143 model .centers = clusters = normalize ? denormalize (task ._cMeans , vecs ) : task ._cMeans ;
141144 model .between_cluster_variances = task ._betwnSqrs ;
142145 double [] variances = new double [task ._cSqrs .length ];
@@ -343,6 +346,7 @@ public static class KMeans2Model extends Model implements Progress {
343346 // Normalization caches
344347 private transient double [][] _normClust ;
345348 private transient double [] _means , _mults ;
349+ private transient int _ncats , _nnums ;
346350
347351 public KMeans2Model (KMeans2 params , Key selfKey , Key dataKey , String names [], String domains [][]) {
348352 super (selfKey , dataKey , names , domains );
@@ -380,7 +384,8 @@ public KMeans2Model(KMeans2 params, Key selfKey, Key dataKey, String names[], St
380384 }
381385 data (tmp , chunks , rowInChunk , _means , _mults );
382386 Arrays .fill (preds , 0 );
383- int cluster = closest (cs , tmp , new ClusterDist ())._cluster ;
387+ // int cluster = closest(cs, tmp, new ClusterDist())._cluster;
388+ int cluster = closest (cs , tmp , _ncats , new ClusterDist ())._cluster ;
384389 preds [0 ] = cluster ; // prediction in preds[0]
385390 preds [1 +cluster ] = 1 ; // class distribution
386391 return preds ;
@@ -401,13 +406,15 @@ public class Clusters extends MRTask2<Clusters> {
401406 // IN
402407 double [][] _clusters ; // Cluster centers
403408 double [] _means , _mults ; // Normalization
409+ int _ncats , _nnums ;
404410
405411 @ Override public void map (Chunk [] cs , NewChunk ncs ) {
406412 double [] values = new double [_clusters [0 ].length ];
407413 ClusterDist cd = new ClusterDist ();
408414 for (int row = 0 ; row < cs [0 ]._len ; row ++) {
409415 data (values , cs , row , _means , _mults );
410- closest (_clusters , values , cd );
416+ // closest(_clusters, values, cd);
417+ closest (_clusters , values , _ncats , cd );
411418 int clu = cd ._cluster ;
412419 // ncs[0].addNum(clu);
413420 ncs .addEnum (clu );
@@ -478,6 +485,7 @@ public static class Lloyds extends MRTask2<Lloyds> {
478485 // IN
479486 double [][] _clusters ;
480487 double [] _means , _mults ; // Normalization
488+ int _ncats , _nnums ;
481489
482490 // OUT
483491 double [][] _cMeans , _cSqrs ; // Means and sum of squares for each cluster
@@ -499,7 +507,8 @@ public static class Lloyds extends MRTask2<Lloyds> {
499507 int [] clusters = new int [cs [0 ]._len ];
500508 for ( int row = 0 ; row < cs [0 ]._len ; row ++ ) {
501509 data (values , cs , row , _means , _mults );
502- closest (_clusters , values , cd );
510+ // closest(_clusters, values, cd);
511+ closest (_clusters , values , _ncats , cd );
503512 int clu = clusters [row ] = cd ._cluster ;
504513 _sqr += cd ._dist ;
505514 if ( clu == -1 )
@@ -556,10 +565,6 @@ private static final class ClusterDist {
556565 double _dist ;
557566 }
558567
559- private static ClusterDist closest (double [][] clusters , double [] point , ClusterDist cd ) {
560- return closest (clusters , point , cd , clusters .length );
561- }
562-
563568 private static double minSqr (double [][] clusters , double [] point , ClusterDist cd ) {
564569 return closest (clusters , point , cd , clusters .length )._dist ;
565570 }
@@ -568,14 +573,43 @@ private static double minSqr(double[][] clusters, double[] point, ClusterDist cd
568573 return closest (clusters , point , cd , count )._dist ;
569574 }
570575
571- /** Return both nearest of N cluster/centroids, and the square-distance. */
576+ private static ClusterDist closest (double [][] clusters , double [] point , ClusterDist cd ) {
577+ return closest (clusters , point , cd , clusters .length );
578+ }
579+
580+ private static ClusterDist closest (double [][] clusters , double [] point , int ncats , ClusterDist cd ) {
581+ return closest (clusters , point , ncats , cd , clusters .length );
582+ }
583+
572584 private static ClusterDist closest (double [][] clusters , double [] point , ClusterDist cd , int count ) {
585+ return closest (clusters , point , 0 , cd , count );
586+ }
587+
588+ private static ClusterDist closest (double [][] clusters , double [] point , int ncats , ClusterDist cd , int count ) {
589+ return closest (clusters , point , ncats , cd , count , 1 );
590+ }
591+
592+ /** Return both nearest of N cluster/centroids, and the square-distance. */
593+ private static ClusterDist closest (double [][] clusters , double [] point , int ncats , ClusterDist cd , int count , double dist ) {
573594 int min = -1 ;
574595 double minSqr = Double .MAX_VALUE ;
575596 for ( int cluster = 0 ; cluster < count ; cluster ++ ) {
576597 double sqr = 0 ; // Sum of dimensional distances
577598 int pts = point .length ; // Count of valid points
578- for ( int column = 0 ; column < clusters [cluster ].length ; column ++ ) {
599+
600+ // Expand categoricals into binary indicator cols
601+ for (int column = 0 ; column < ncats ; column ++) {
602+ double d = point [column ];
603+ if (Double .isNaN (d ))
604+ pts --;
605+ else {
606+ // TODO: What is the distance between unequal categoricals?
607+ if (d != clusters [cluster ][column ])
608+ sqr += 2 * dist * dist ;
609+ }
610+ }
611+
612+ for ( int column = ncats ; column < clusters [cluster ].length ; column ++ ) {
579613 double d = point [column ];
580614 if ( Double .isNaN (d ) ) { // Bad data?
581615 pts --; // Do not count
@@ -686,14 +720,16 @@ private static double[][] denormalize(double[][] clusters, Vec[] vecs) {
686720 private static void data (double [] values , Vec [] vecs , long row , double [] means , double [] mults ) {
687721 for ( int i = 0 ; i < values .length ; i ++ ) {
688722 double d = vecs [i ].at (row );
689- values [i ] = data (d , i , means , mults );
723+ // values[i] = data(d, i, means, mults);
724+ values [i ] = data (d , i , means , mults , vecs [i ].cardinality ());
690725 }
691726 }
692727
693728 private static void data (double [] values , Chunk [] chks , int row , double [] means , double [] mults ) {
694729 for ( int i = 0 ; i < values .length ; i ++ ) {
695730 double d = chks [i ].at0 (row );
696- values [i ] = data (d , i , means , mults );
731+ // values[i] = data(d, i, means, mults);
732+ values [i ] = data (d , i , means , mults , chks [i ]._vec .cardinality ());
697733 }
698734 }
699735
@@ -709,4 +745,20 @@ private static double data(double d, int i, double[] means, double[] mults) {
709745 }
710746 return d ;
711747 }
748+
749+ private static double data (double d , int i , double [] means , double [] mults , int cardinality ) {
750+ if (cardinality == -1 ) {
751+ if ( Double .isNaN (d ) )
752+ d = means [i ];
753+ if ( mults != null ) {
754+ d -= means [i ];
755+ d *= mults [i ];
756+ }
757+ } else {
758+ // TODO: If NaN, then replace with majority class?
759+ if (Double .isNaN (d ))
760+ d = Math .min (Math .round (means [i ]), cardinality -1 );
761+ }
762+ return d ;
763+ }
712764}
0 commit comments