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.timeseries.aggregator; 031 032import java.util.ArrayList; 033import java.util.Iterator; 034import java.util.List; 035import java.util.Set; 036 037import org.openimaj.ml.regression.LinearRegression; 038import org.openimaj.ml.timeseries.collection.SynchronisedTimeSeriesCollection; 039import org.openimaj.ml.timeseries.series.DoubleSynchronisedTimeSeriesCollection; 040import org.openimaj.ml.timeseries.series.DoubleTimeSeries; 041import org.openimaj.util.pair.IndependentPair; 042 043/** 044 * An implementation of a general linear regressive such that the values of a timeseries Y are predicted using 045 * the values of a set of time series X at some offset over some time window. X may potentially contain Y itself 046 * which turns this into an auto-regressive model augmented with extra information. Furthermore, varying window 047 * sizes and offsets may be used for each time series X. 048 * 049 * This is all achieved with {@link SynchronisedTimeSeriesCollection} which model a set of timeseries which are 050 * synchronised. 051 * 052 * When intitalised, the Y time series must be explicitly specified. By default the 053 * @author Sina Samangooei (ss@ecs.soton.ac.uk) 054 * 055 */ 056public class WindowedLinearRegressionAggregator implements SynchronisedTimeSeriesCollectionAggregator< 057 DoubleTimeSeries, 058 DoubleSynchronisedTimeSeriesCollection, 059 DoubleTimeSeries>{ 060 061 private static final int DEFAULT_WINDOW_SIZE = 3; 062 private static final int DEFAULT_OFFSET = 1; 063 private LinearRegression reg; 064 private boolean autoregressive = true; 065 private List<IndependentPair<Integer,Integer>> windowOffsets; 066 private String ydataName; 067 068 private WindowedLinearRegressionAggregator(){ 069 this.windowOffsets = new ArrayList<IndependentPair<Integer,Integer>>(); 070 } 071 /** 072 * Calculate the regression from the same time series inputed 073 * @param ydataName 074 */ 075 public WindowedLinearRegressionAggregator(String ydataName) { 076 this(); 077 windowOffsets.add(IndependentPair.pair(DEFAULT_WINDOW_SIZE, DEFAULT_OFFSET)); 078 this.ydataName = ydataName; 079 } 080 081 /** 082 * Calculate the regression from the same time series inputed 083 * @param ydataName 084 * @param autoregressive whether the ydata should be used in regression 085 */ 086 public WindowedLinearRegressionAggregator(String ydataName,boolean autoregressive) { 087 this(); 088 windowOffsets.add(IndependentPair.pair(DEFAULT_WINDOW_SIZE, DEFAULT_OFFSET)); 089 this.ydataName = ydataName; 090 this.autoregressive = autoregressive; 091 } 092 093 /** 094 * Perform regression s.t. Y = Sum(w_{0-i} * x_{0-i}) + c using the same window size 095 * for all other time series 096 * @param ydataName 097 * @param windowsize 098 */ 099 public WindowedLinearRegressionAggregator(String ydataName,int windowsize) { 100 this(); 101 this.ydataName = ydataName; 102 windowOffsets.add(IndependentPair.pair(windowsize, DEFAULT_OFFSET)); 103 104 } 105 /** 106 * Perform regression s.t. Y = Sum(w_{0-i} * x_{0-i}) + c using the same window size 107 * for all other time series 108 * @param ydataName 109 * @param windowsize 110 * @param autoregressive whether the ydata should be used in regression 111 */ 112 public WindowedLinearRegressionAggregator(String ydataName,int windowsize,boolean autoregressive) { 113 this(); 114 this.ydataName = ydataName; 115 windowOffsets.add(IndependentPair.pair(windowsize, DEFAULT_OFFSET)); 116 this.autoregressive = autoregressive; 117 118 } 119 120 /** 121 * Perform regression s.t. y = Sum(w_{0-i} * x_{0-i}) + c for i from 1 to windowsize with some 122 * offset. The same windowsize and offset is used for each time series 123 * @param ydataName 124 * @param windowsize 125 * @param offset 126 */ 127 public WindowedLinearRegressionAggregator(String ydataName,int windowsize, int offset) { 128 this(); 129 this.ydataName = ydataName; 130 windowOffsets.add(IndependentPair.pair(windowsize, offset)); 131 132 } 133 134 /** 135 * Perform regression s.t. y = Sum(w_{0-i} * x_{0-i}) + c for i from 1 to windowsize with some 136 * offset. The same windowsize and offset is used for each time series 137 * @param ydataName 138 * @param windowsize 139 * @param offset 140 * @param autoregressive whether the ydata should be used in regression 141 */ 142 public WindowedLinearRegressionAggregator(String ydataName,int windowsize, int offset, boolean autoregressive) { 143 this(); 144 this.ydataName = ydataName; 145 this.autoregressive = autoregressive; 146 windowOffsets.add(IndependentPair.pair(windowsize, offset)); 147 148 } 149 150 /** 151 * Perform regression s.t. y = Sum(w_{0-i} * x_{0-i}) + c for i from 1 to windowsize with some 152 * offset. The same windowsize and offset is used for each time series 153 * @param ydataName 154 * @param windowsize 155 * @param offset 156 * @param autoregressive whether the ydata should be used in regression 157 */ 158 public WindowedLinearRegressionAggregator(String ydataName,int windowsize, int offset, boolean autoregressive,DoubleSynchronisedTimeSeriesCollection other) { 159 this(); 160 WindowedLinearRegressionAggregator regress = new WindowedLinearRegressionAggregator(ydataName,windowsize,offset,autoregressive); 161 regress.aggregate(other); 162 this.reg =regress.reg; 163 this.ydataName = regress.ydataName; 164 this.autoregressive = regress.autoregressive; 165 this.windowOffsets = regress.windowOffsets; 166 } 167 168 /** 169 * Perform regression s.t. y = Sum(w_{0-i} * x_{0-i}) + c for i from 1 to windowsize with some 170 * offset. The same windowsize and offset is used for each time series 171 * @param ydataName 172 * @param autoregressive whether the ydata should be used in regression 173 * @param windowOffsets 174 */ 175 public WindowedLinearRegressionAggregator(String ydataName,boolean autoregressive,IndependentPair<Integer,Integer> ... windowOffsets) { 176 this(); 177 this.ydataName = ydataName; 178 this.autoregressive = autoregressive; 179 for (IndependentPair<Integer, Integer> independentPair : windowOffsets) { 180 this.windowOffsets.add(independentPair); 181 } 182 } 183 184 185// @Override 186// public void process(DoubleTimeSeries series) { 187// Matrix x = new Matrix(new double[][]{ArrayUtils.longToDouble(series.getTimes())}).transpose(); 188// List<IndependentPair<double[], double[]>> instances = new ArrayList<IndependentPair<double[], double[]>>(); 189// double[] data = series.getData(); 190// 191// for (int i = this.windowsize + (offset - 1); i < series.size(); i++) { 192// int start = i - this.windowsize - (offset - 1); 193//// System.out.format("Range %d->%d (inclusive) used to calculate: %d\n",start,start+this.windowsize-1,i); 194// double[] datawindow = new double[this.windowsize]; 195// System.arraycopy(data, start, datawindow, 0, this.windowsize); 196// instances.add(IndependentPair.pair(datawindow, new double[]{data[i]})); 197// } 198// if(!regdefined) 199// { 200// this.reg = new LinearRegression(); 201// this.reg.estimate(instances); 202// } 203// System.out.println(this.reg); 204// Iterator<IndependentPair<double[], double[]>> instanceIter = instances.iterator(); 205// for (int i = this.windowsize + (offset - 1); i < series.size(); i++) { 206// data[i] = this.reg.predict(instanceIter.next().firstObject())[0]; 207// } 208// } 209 210 /** 211 * @return the {@link WindowedLinearRegressionAggregator}'s underlying {@link LinearRegression} model 212 */ 213 public LinearRegression getRegression() { 214 return this.reg; 215 } 216 217 @Override 218 public DoubleTimeSeries aggregate(DoubleSynchronisedTimeSeriesCollection series) { 219 220 Set<String> names = series.getNames(); 221 if(!autoregressive){ 222 names.remove(ydataName); 223 } 224 DoubleTimeSeries yseries = series.series(ydataName ); 225 double[] ydata = yseries.getData(); 226 227 series = series.collectionByNames(names); 228 double[] data = series.flatten(); 229 if(this.windowOffsets.size() != series.nSeries() && this.windowOffsets.size() == 1){ 230 IndependentPair<Integer, Integer> offset = this.windowOffsets.get(0); 231 return aggregteSingle(yseries.getTimes(),ydata,data,offset.firstObject(),offset.secondObject(),series.nSeries()); 232 } 233 234 return null; 235 } 236 237 private DoubleTimeSeries aggregteSingle(long[] times, double[] ydata, double[] data,int windowsize, int offset, int nseries) { 238 List<IndependentPair<double[], double[]>> instances = new ArrayList<IndependentPair<double[], double[]>>(); 239 for (int i = windowsize + (offset - 1); i < ydata.length; i++) { 240 int start = (i - windowsize - (offset - 1)) * nseries; 241// System.out.format("Range %d->%d (inclusive) used to calculate: %d\n",start,start+this.windowsize-1,i); 242 double[] datawindow = new double[windowsize*nseries]; 243 System.arraycopy(data, start, datawindow, 0, windowsize*nseries); 244 instances.add(IndependentPair.pair(datawindow, new double[]{ydata[i]})); 245 } 246 if(this.reg==null){ 247 this.reg = new LinearRegression(); 248 this.reg.estimate(instances); 249 } 250 251 DoubleTimeSeries ret = new DoubleTimeSeries(times,new double[ydata.length]); 252 253 Iterator<IndependentPair<double[], double[]>> instanceIter = instances.iterator(); 254 data = ret.getData(); 255 for (int i = windowsize + (offset - 1); i < ydata.length; i++) { 256 double[] predicted = this.reg.predict(instanceIter.next().firstObject()); 257 data[i] = predicted[0]; 258 } 259 return ret.get(times[windowsize + (offset-1)], times[ydata.length-1]); 260 } 261 262 public LinearRegression getReg(){ 263 return this.reg; 264 } 265 266}