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 045 * timeseries Y are predicted using the values of a set of time series X at some 046 * offset over some time window. X may potentially contain Y itself which turns 047 * this into an auto-regressive model augmented with extra information. 048 * Furthermore, varying window sizes and offsets may be used for each time 049 * series X. 050 * 051 * This is all achieved with {@link SynchronisedTimeSeriesCollection} which 052 * model a set of timeseries which are synchronised. 053 * 054 * When intitalised, the Y time series must be explicitly specified. By default 055 * the 056 * 057 * @author Sina Samangooei (ss@ecs.soton.ac.uk) 058 * 059 */ 060public class WindowedLinearRegressionAggregator implements SynchronisedTimeSeriesCollectionAggregator< 061 DoubleTimeSeries, 062 DoubleSynchronisedTimeSeriesCollection, 063 DoubleTimeSeries> 064{ 065 066 private static final int DEFAULT_WINDOW_SIZE = 3; 067 private static final int DEFAULT_OFFSET = 1; 068 private LinearRegression reg; 069 private boolean autoregressive = true; 070 private List<IndependentPair<Integer, Integer>> windowOffsets; 071 private String ydataName; 072 073 private WindowedLinearRegressionAggregator() { 074 this.windowOffsets = new ArrayList<IndependentPair<Integer, Integer>>(); 075 } 076 077 /** 078 * Calculate the regression from the same time series inputed 079 * 080 * @param ydataName 081 */ 082 public WindowedLinearRegressionAggregator(String ydataName) { 083 this(); 084 windowOffsets.add(IndependentPair.pair(DEFAULT_WINDOW_SIZE, DEFAULT_OFFSET)); 085 this.ydataName = ydataName; 086 } 087 088 /** 089 * Calculate the regression from the same time series inputed 090 * 091 * @param ydataName 092 * @param autoregressive 093 * whether the ydata should be used in regression 094 */ 095 public WindowedLinearRegressionAggregator(String ydataName, boolean autoregressive) { 096 this(); 097 windowOffsets.add(IndependentPair.pair(DEFAULT_WINDOW_SIZE, DEFAULT_OFFSET)); 098 this.ydataName = ydataName; 099 this.autoregressive = autoregressive; 100 } 101 102 /** 103 * Perform regression s.t. Y = Sum(w_{0-i} * x_{0-i}) + c using the same 104 * window size for all other time series 105 * 106 * @param ydataName 107 * @param windowsize 108 */ 109 public WindowedLinearRegressionAggregator(String ydataName, int windowsize) { 110 this(); 111 this.ydataName = ydataName; 112 windowOffsets.add(IndependentPair.pair(windowsize, DEFAULT_OFFSET)); 113 114 } 115 116 /** 117 * Perform regression s.t. Y = Sum(w_{0-i} * x_{0-i}) + c using the same 118 * window size for all other time series 119 * 120 * @param ydataName 121 * @param windowsize 122 * @param autoregressive 123 * whether the ydata should be used in regression 124 */ 125 public WindowedLinearRegressionAggregator(String ydataName, int windowsize, boolean autoregressive) { 126 this(); 127 this.ydataName = ydataName; 128 windowOffsets.add(IndependentPair.pair(windowsize, DEFAULT_OFFSET)); 129 this.autoregressive = autoregressive; 130 131 } 132 133 /** 134 * Perform regression s.t. y = Sum(w_{0-i} * x_{0-i}) + c for i from 1 to 135 * windowsize with some offset. The same windowsize and offset is used for 136 * each time series 137 * 138 * @param ydataName 139 * @param windowsize 140 * @param offset 141 */ 142 public WindowedLinearRegressionAggregator(String ydataName, int windowsize, int offset) { 143 this(); 144 this.ydataName = ydataName; 145 windowOffsets.add(IndependentPair.pair(windowsize, offset)); 146 147 } 148 149 /** 150 * Perform regression s.t. y = Sum(w_{0-i} * x_{0-i}) + c for i from 1 to 151 * windowsize with some offset. The same windowsize and offset is used for 152 * each time series 153 * 154 * @param ydataName 155 * @param windowsize 156 * @param offset 157 * @param autoregressive 158 * whether the ydata should be used in regression 159 */ 160 public WindowedLinearRegressionAggregator(String ydataName, int windowsize, int offset, boolean autoregressive) { 161 this(); 162 this.ydataName = ydataName; 163 this.autoregressive = autoregressive; 164 windowOffsets.add(IndependentPair.pair(windowsize, offset)); 165 166 } 167 168 /** 169 * Perform regression s.t. y = Sum(w_{0-i} * x_{0-i}) + c for i from 1 to 170 * windowsize with some offset. The same windowsize and offset is used for 171 * each time series 172 * 173 * @param ydataName 174 * @param windowsize 175 * @param offset 176 * @param autoregressive 177 * whether the ydata should be used in regression 178 * @param other 179 */ 180 public WindowedLinearRegressionAggregator(String ydataName, int windowsize, int offset, boolean autoregressive, 181 DoubleSynchronisedTimeSeriesCollection other) 182 { 183 this(); 184 final WindowedLinearRegressionAggregator regress = new WindowedLinearRegressionAggregator(ydataName, windowsize, 185 offset, autoregressive); 186 regress.aggregate(other); 187 this.reg = regress.reg; 188 this.ydataName = regress.ydataName; 189 this.autoregressive = regress.autoregressive; 190 this.windowOffsets = regress.windowOffsets; 191 } 192 193 /** 194 * Perform regression s.t. y = Sum(w_{0-i} * x_{0-i}) + c for i from 1 to 195 * windowsize with some offset. The same windowsize and offset is used for 196 * each time series 197 * 198 * @param ydataName 199 * @param autoregressive 200 * whether the ydata should be used in regression 201 * @param windowOffsets 202 */ 203 @SafeVarargs 204 public WindowedLinearRegressionAggregator(String ydataName, boolean autoregressive, 205 IndependentPair<Integer, Integer>... windowOffsets) 206 { 207 this(); 208 this.ydataName = ydataName; 209 this.autoregressive = autoregressive; 210 for (final IndependentPair<Integer, Integer> independentPair : windowOffsets) { 211 this.windowOffsets.add(independentPair); 212 } 213 } 214 215 // @Override 216 // public void process(DoubleTimeSeries series) { 217 // Matrix x = new Matrix(new 218 // double[][]{ArrayUtils.longToDouble(series.getTimes())}).transpose(); 219 // List<IndependentPair<double[], double[]>> instances = new 220 // ArrayList<IndependentPair<double[], double[]>>(); 221 // double[] data = series.getData(); 222 // 223 // for (int i = this.windowsize + (offset - 1); i < series.size(); i++) { 224 // int start = i - this.windowsize - (offset - 1); 225 // // 226 // System.out.format("Range %d->%d (inclusive) used to calculate: %d\n",start,start+this.windowsize-1,i); 227 // double[] datawindow = new double[this.windowsize]; 228 // System.arraycopy(data, start, datawindow, 0, this.windowsize); 229 // instances.add(IndependentPair.pair(datawindow, new double[]{data[i]})); 230 // } 231 // if(!regdefined) 232 // { 233 // this.reg = new LinearRegression(); 234 // this.reg.estimate(instances); 235 // } 236 // System.out.println(this.reg); 237 // Iterator<IndependentPair<double[], double[]>> instanceIter = 238 // instances.iterator(); 239 // for (int i = this.windowsize + (offset - 1); i < series.size(); i++) { 240 // data[i] = this.reg.predict(instanceIter.next().firstObject())[0]; 241 // } 242 // } 243 244 /** 245 * @return the {@link WindowedLinearRegressionAggregator}'s underlying 246 * {@link LinearRegression} model 247 */ 248 public LinearRegression getRegression() { 249 return this.reg; 250 } 251 252 @Override 253 public DoubleTimeSeries aggregate(DoubleSynchronisedTimeSeriesCollection series) { 254 255 final Set<String> names = series.getNames(); 256 if (!autoregressive) { 257 names.remove(ydataName); 258 } 259 final DoubleTimeSeries yseries = series.series(ydataName); 260 final double[] ydata = yseries.getData(); 261 262 series = series.collectionByNames(names); 263 final double[] data = series.flatten(); 264 if (this.windowOffsets.size() != series.nSeries() && this.windowOffsets.size() == 1) { 265 final IndependentPair<Integer, Integer> offset = this.windowOffsets.get(0); 266 return aggregteSingle(yseries.getTimes(), ydata, data, offset.firstObject(), offset.secondObject(), 267 series.nSeries()); 268 } 269 270 return null; 271 } 272 273 private DoubleTimeSeries aggregteSingle(long[] times, double[] ydata, double[] data, int windowsize, int offset, 274 int nseries) 275 { 276 final List<IndependentPair<double[], double[]>> instances = new ArrayList<IndependentPair<double[], double[]>>(); 277 for (int i = windowsize + (offset - 1); i < ydata.length; i++) { 278 final int start = (i - windowsize - (offset - 1)) * nseries; 279 // System.out.format("Range %d->%d (inclusive) used to calculate: %d\n",start,start+this.windowsize-1,i); 280 final double[] datawindow = new double[windowsize * nseries]; 281 System.arraycopy(data, start, datawindow, 0, windowsize * nseries); 282 instances.add(IndependentPair.pair(datawindow, new double[] { ydata[i] })); 283 } 284 if (this.reg == null) { 285 this.reg = new LinearRegression(); 286 this.reg.estimate(instances); 287 } 288 289 final DoubleTimeSeries ret = new DoubleTimeSeries(times, new double[ydata.length]); 290 291 final Iterator<IndependentPair<double[], double[]>> instanceIter = instances.iterator(); 292 data = ret.getData(); 293 for (int i = windowsize + (offset - 1); i < ydata.length; i++) { 294 final double[] predicted = this.reg.predict(instanceIter.next().firstObject()); 295 data[i] = predicted[0]; 296 } 297 return ret.get(times[windowsize + (offset - 1)], times[ydata.length - 1]); 298 } 299 300 public LinearRegression getReg() { 301 return this.reg; 302 } 303 304}