001/**
002 * Copyright (c) 2011, The University of Southampton and the individual contributors.
003 * All rights reserved.
004 *
005 * Redistribution and use in source and binary forms, with or without modification,
006 * are permitted provided that the following conditions are met:
007 *
008 *   *  Redistributions of source code must retain the above copyright notice,
009 *      this list of conditions and the following disclaimer.
010 *
011 *   *  Redistributions in binary form must reproduce the above copyright notice,
012 *      this list of conditions and the following disclaimer in the documentation
013 *      and/or other materials provided with the distribution.
014 *
015 *   *  Neither the name of the University of Southampton nor the names of its
016 *      contributors may be used to endorse or promote products derived from this
017 *      software without specific prior written permission.
018 *
019 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
020 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
021 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
022 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
023 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
024 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
025 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
026 * ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
027 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
028 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
029 */
030package org.openimaj.ml.clustering.spectral;
031
032import java.util.Iterator;
033
034import org.openimaj.util.pair.DoubleObjectPair;
035
036import ch.akuhn.matrix.SparseMatrix;
037import ch.akuhn.matrix.Vector;
038import ch.akuhn.matrix.eigenvalues.FewEigenvalues;
039
040/**
041 * Attempts to automatically choose the number of eigen vectors based on the
042 * relative gap between eigen values. In spectral clustering the gap between the
043 * eigen values of "good" clusters jumps. This class ignores the gap between 0 and
044 * the next item because 0s represent completely isolated objects and in all but the trivial
045 * case we must stop after we have run out of 0s.
046 * @author Sina Samangooei (ss@ecs.soton.ac.uk)
047 *
048 */
049public class ChangeDetectingEigenChooser extends EigenChooser {
050
051        private double relativeGap;
052        private double maxSelect;
053
054        /**
055         * @param relativeGap the gap between previous and current (treated as absolute if previous value == 0)
056         * @param maxSelect
057         */
058        public ChangeDetectingEigenChooser(double relativeGap, double maxSelect) {
059                this.relativeGap = relativeGap;
060                this.maxSelect = maxSelect;
061        }
062
063        @Override
064        public int nEigenVectors(Iterator<DoubleObjectPair<Vector>> vals, int totalEigenVectors) {
065                int count = 0;
066                double prevDiff = 0;
067                double prevVal = vals.next().first;
068                for (;vals.hasNext();) {
069                        double val = vals.next().first;
070                        if(val < 0) break;
071                        double diff = Math.abs(val - prevVal);
072                        if(prevDiff != 0){
073                                double l = prevDiff * relativeGap;
074                                if(diff > l) {
075                                        count++;
076                                        break;
077                                }
078                        }
079                        prevDiff = diff;
080                        prevVal = val;
081                        count ++;
082                }
083                int maxCount = (int) (totalEigenVectors * maxSelect);
084                if(count > maxCount){
085                        return maxCount;
086                }
087                return count;
088        }
089
090        @Override
091        public FewEigenvalues prepare(final SparseMatrix laplacian) {
092                int total = laplacian.columnCount();
093                FewEigenvalues eig = FewEigenvalues.of(laplacian);
094                return eig.greatest((int) (total*maxSelect));
095        }
096
097
098}