@@ -64,18 +64,21 @@ public class ConditionalSampleSummarizer {
6464 */
6565 protected boolean project = false ;
6666
67- public ConditionalSampleSummarizer (int [] missingDimensions , float [] queryPoint , double centrality ) {
68- this .missingDimensions = Arrays .copyOf (missingDimensions , missingDimensions .length );
69- this .queryPoint = Arrays .copyOf (queryPoint , queryPoint .length );
70- this .centrality = centrality ;
71- }
67+ protected int numberOfReps = 1 ;
7268
73- public ConditionalSampleSummarizer (int [] missingDimensions , float [] queryPoint , double centrality ,
74- boolean project ) {
69+ protected double shrinkage = 0 ;
70+
71+ protected int shingleSize = 1 ;
72+
73+ public ConditionalSampleSummarizer (int [] missingDimensions , float [] queryPoint , double centrality , boolean project ,
74+ int numberOfReps , double shrinkage , int shingleSize ) {
7575 this .missingDimensions = Arrays .copyOf (missingDimensions , missingDimensions .length );
7676 this .queryPoint = Arrays .copyOf (queryPoint , queryPoint .length );
7777 this .centrality = centrality ;
7878 this .project = project ;
79+ this .numberOfReps = numberOfReps ;
80+ this .shrinkage = shrinkage ;
81+ this .shingleSize = shingleSize ;
7982 }
8083
8184 public SampleSummary summarize (List <ConditionalTreeSample > alist ) {
@@ -102,21 +105,28 @@ public SampleSummary summarize(List<ConditionalTreeSample> alist, boolean addTyp
102105 List <ConditionalTreeSample > newList = ConditionalTreeSample .dedup (alist );
103106
104107 newList .sort ((o1 , o2 ) -> Double .compare (o1 .distance , o2 .distance ));
108+ int dimensions = queryPoint .length ;
105109
106- ArrayList <Weighted <float []>> points = new ArrayList <>();
107- newList .stream ().forEach (e -> {
108- if (!project ) {
109- points .add (new Weighted <>(e .leafPoint , (float ) e .weight ));
110- } else {
111- float [] values = new float [missingDimensions .length ];
112- for (int i = 0 ; i < missingDimensions .length ; i ++) {
113- values [i ] = e .leafPoint [missingDimensions [i ]];
110+ if (!addTypical ) {
111+ ArrayList <Weighted <float []>> points = new ArrayList <>();
112+ newList .stream ().forEach (e -> {
113+ if (!project ) {
114+ if (shingleSize == 1 ) {
115+ points .add (new Weighted <>(e .leafPoint , (float ) e .weight ));
116+ } else {
117+ float [] values = Arrays .copyOfRange (e .leafPoint , dimensions - dimensions / shingleSize ,
118+ dimensions );
119+ points .add (new Weighted <>(values , (float ) e .weight ));
120+ }
121+ } else {
122+ float [] values = new float [missingDimensions .length ];
123+ for (int i = 0 ; i < missingDimensions .length ; i ++) {
124+ values [i ] = e .leafPoint [missingDimensions [i ]];
125+ }
126+ points .add (new Weighted <>(values , (float ) e .weight ));
114127 }
115- points .add (new Weighted <>(values , (float ) e .weight ));
116- }
117- });
128+ });
118129
119- if (!addTypical ) {
120130 return new SampleSummary (points );
121131 }
122132
@@ -131,34 +141,37 @@ public SampleSummary summarize(List<ConditionalTreeSample> alist, boolean addTyp
131141 * exact matches would go against the dynamic sampling based use of RCF.
132142 **/
133143
134- int dimensions = queryPoint .length ;
135-
136- double threshold = centrality * newList .get (0 ).distance ;
137- double currentWeight = 0 ;
138- int alwaysInclude = 0 ;
139- double remainderWeight = totalWeight ;
140- while (newList .get (alwaysInclude ).distance == 0 ) {
141- remainderWeight -= newList .get (alwaysInclude ).weight ;
142- ++alwaysInclude ;
143- if (alwaysInclude == newList .size ()) {
144- break ;
144+ int num = 0 ;
145+ if (centrality > 0 ) {
146+ double threshold = centrality * newList .get (0 ).distance + 1e-6 ;
147+ double currentWeight = 0 ;
148+ int alwaysInclude = 0 ;
149+ double remainderWeight = totalWeight ;
150+ while (newList .get (alwaysInclude ).distance == 0 ) {
151+ remainderWeight -= newList .get (alwaysInclude ).weight ;
152+ ++alwaysInclude ;
153+ if (alwaysInclude == newList .size ()) {
154+ break ;
155+ }
145156 }
146- }
147- for (int j = 1 ; j < newList .size (); j ++) {
148- if ((currentWeight < remainderWeight / 3 && currentWeight + newList .get (j ).weight >= remainderWeight / 3 )
149- || (currentWeight < remainderWeight / 2
150- && currentWeight + newList .get (j ).weight >= remainderWeight / 2 )) {
151- threshold = centrality * newList .get (j ).distance ;
157+ for (int j = 1 ; j < newList .size (); j ++) {
158+ if ((currentWeight < remainderWeight / 3
159+ && currentWeight + newList .get (j ).weight >= remainderWeight / 3 )
160+ || (currentWeight < remainderWeight / 2
161+ && currentWeight + newList .get (j ).weight >= remainderWeight / 2 )) {
162+ threshold = centrality * newList .get (j ).distance ;
163+ }
164+ currentWeight += newList .get (j ).weight ;
152165 }
153- currentWeight += newList . get ( j ). weight ;
154- }
155- // note that the threshold is currently centrality * (some distance in the list)
156- // thus the sequel uses a convex combination; and setting centrality = 0 removes
157- // the entire filtering based on distances
158- threshold += ( 1 - centrality ) * newList . get ( newList . size () - 1 ). distance ;
159- int num = 0 ;
160- while ( num < newList . size () && newList . get ( num ). distance <= threshold ) {
161- ++ num ;
166+ // note that the threshold is currently centrality * (some distance in the list)
167+ // thus the sequel uses a convex combination; and setting centrality = 0 removes
168+ // the entire filtering based on distances
169+ threshold += ( 1 - centrality ) * newList . get ( newList . size () - 1 ). distance ;
170+ while ( num < newList . size () && newList . get ( num ). distance <= threshold ) {
171+ ++ num ;
172+ }
173+ } else {
174+ num = newList . size () ;
162175 }
163176
164177 ArrayList <Weighted <float []>> typicalPoints = new ArrayList <>();
@@ -171,26 +184,21 @@ public SampleSummary summarize(List<ConditionalTreeSample> alist, boolean addTyp
171184 values [i ] = e .leafPoint [missingDimensions [i ]];
172185 }
173186 } else {
174- values = Arrays .copyOf (e .leafPoint , dimensions );
187+ if (shingleSize == 1 ) {
188+ values = e .leafPoint ;
189+ } else {
190+ values = Arrays .copyOfRange (e .leafPoint , dimensions - dimensions / shingleSize , dimensions );
191+ }
175192 }
176193 typicalPoints .add (new Weighted <>(values , (float ) e .weight ));
177194 }
178195 int maxAllowed = min (queryPoint .length * MAX_NUMBER_OF_TYPICAL_PER_DIMENSION , MAX_NUMBER_OF_TYPICAL_ELEMENTS );
179196 maxAllowed = min (maxAllowed , num );
180- SampleSummary projectedSummary = Summarizer .l2summarize (typicalPoints , maxAllowed , num , false , 72 );
181197
182- float [][] pointList = new float [projectedSummary .summaryPoints .length ][];
183- float [] likelihood = new float [projectedSummary .summaryPoints .length ];
184-
185- for (int i = 0 ; i < projectedSummary .summaryPoints .length ; i ++) {
186- pointList [i ] = Arrays .copyOf (queryPoint , dimensions );
187- for (int j = 0 ; j < missingDimensions .length ; j ++) {
188- pointList [i ][missingDimensions [j ]] = projectedSummary .summaryPoints [i ][j ];
189- }
190- likelihood [i ] = projectedSummary .relativeWeight [i ];
191- }
198+ SampleSummary projectedSummary = Summarizer .summarize (typicalPoints , maxAllowed , num , false ,
199+ Summarizer ::L2distance , 72 , false , numberOfReps , shrinkage );
192200
193- return new SampleSummary (points , pointList , likelihood );
201+ return new SampleSummary (typicalPoints , projectedSummary );
194202 }
195203
196204}
0 commit comments