/*
* Copyright (c) 2010 Pentaho Corporation. All rights reserved.
* This software was developed by Pentaho Corporation and is provided under the terms
* of the GNU Lesser General Public License, Version 2.1. You may not use
* this file except in compliance with the license. If you need a copy of the license,
* please go to http://www.gnu.org/licenses/lgpl-2.1.txt. The Original Code is Time Series
* Forecasting. The Initial Developer is Pentaho Corporation.
*
* Software distributed under the GNU Lesser Public License is distributed on an "AS IS"
* basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. Please refer to
* the license for the specific language governing your rights and limitations.
*/
/*
* RRSEModule.java
* Copyright (C) 2010 Pentaho Corporation
*/
package weka.classifiers.timeseries.eval;
import java.util.List;
import weka.classifiers.evaluation.NumericPrediction;
import weka.core.Instance;
import weka.core.Utils;
/**
* An evaluation module that computes the root relative squared error
* of forecasted values. I.e. the root mean squared error of forecasted values
* is computed by this module and these are divided by the root mean squared
* error obtained by using a target value from a previous time step
* as the predicted value.
*
* @author Mark Hall (mhall{[at]}pentaho{[dot]}com)
* @version $Revision: 49983 $
*/
public class RRSEModule extends ErrorModule {
protected double[] m_previousActual;
protected double[] m_sumOfSE;
/**
* Holds the RRSE module that this one is relative to - i.e.
* computations of the predictions provided to this module
* will be relative to the actual target values obtained from
* m_relativeRRSE.getPreviousActual(). If no previous RRSEModule
* is set, then this module will use immediately previous actual
* values accumulated as evaluateForInstance() is called (i.e.
* evaluation is relative to using the immediately preceding
* actual value as the forecast. Setting a previous RRSEModule
* allows evaluation relative to actual values further back in
* the past
*/
protected RRSEModule m_relativeRRSE;
protected static final double SMALL = 0.000001;
/**
* Reset this module
*/
public void reset() {
super.reset();
m_previousActual = new double[m_targetFieldNames.size()];
m_sumOfSE = new double[m_targetFieldNames.size()];
for (int i = 0; i < m_targetFieldNames.size(); i++) {
m_previousActual[i] = Utils.missingValue();
m_sumOfSE[i] = 0;
}
}
/**
* Set a RRSEModule to use for the relative calculations - i.e.
* actual target values from this module will be used.
*
* @param relative the RRSE module to use for relative computations.
*/
public void setRelativeRRSEModule(RRSEModule relative) {
m_relativeRRSE = relative;
}
/**
* Get the actual target values from the immediately preceding
* time step.
*
* @return the actual target values from the immediately preceding
* time step.
*/
public double[] getPreviousActual() {
return m_previousActual;
}
/**
* Return the short identifying name of this evaluation module
*
* @return the short identifying name of this evaluation module
*/
public String getEvalName() {
return "RRSE";
}
/**
* Return the longer (single sentence) description
* of this evaluation module
*
* @return the longer description of this module
*/
public String getDescription() {
return "Root relative squared error";
}
/**
* Return the mathematical formula that this
* evaluation module computes.
*
* @return the mathematical formula that this module
* computes.
*/
public String getDefinition() {
return "sqrt(sum((predicted - actual)^2) / N) / " +
"sqrt(sum(previous_target - actual)^2) / N)";
}
protected void evaluatePredictionForTargetForInstance(int targetIndex,
NumericPrediction forecast, double actualValue) {
double predictedValue = forecast.predicted();
double[][] intervals = forecast.predictionIntervals();
NumericPrediction pred = new NumericPrediction(actualValue, predictedValue, 1,
intervals);
m_predictions.get(targetIndex).add(pred);
m_counts[targetIndex]++;
}
/**
* Evaluate the given forecast(s) with respect to the given
* test instance. Targets with missing values are ignored.
*
* @param forecasts a List of forecasted values. Each element
* corresponds to one of the targets and is assumed to be in the same
* order as the list of targets supplied to the setTargetFields() method.
* @throws Exception if the evaluation can't be completed for some
* reason.
*/
public void evaluateForInstance(List<NumericPrediction> forecasts,
Instance inst) throws Exception {
// here just compute the running sum of abs errors for each target
// with respect to using the previous value of the target as a prediction
for (int i = 0; i < m_targetFieldNames.size(); i++) {
double actualValue = getTargetValue(m_targetFieldNames.get(i), inst);
if (m_relativeRRSE != null) {
m_previousActual = m_relativeRRSE.getPreviousActual();
}
if (m_relativeRRSE == null &&
Utils.isMissingValue(m_previousActual[i])) {
m_previousActual[i] = actualValue;
} else {
// only compute for non-missing previous actual values and non-missing
// current actual values
if (!Utils.isMissingValue(actualValue) &&
!Utils.isMissingValue(m_previousActual[i])) {
evaluatePredictionForTargetForInstance(i, forecasts.get(i), actualValue);
m_sumOfSE[i] += ((m_previousActual[i] - actualValue) *
(m_previousActual[i] - actualValue));
// m_newCounts[i]++;
}
if (m_relativeRRSE == null) {
m_previousActual[i] = actualValue;
}
}
}
}
public double[] calculateMeasure() throws Exception {
double[] result = new double[m_targetFieldNames.size()];
for (int i = 0; i < result.length; i++) {
result[i] = Utils.missingValue();
}
for (int i = 0; i < m_targetFieldNames.size(); i++) {
double sumSE = 0;
double count = 0;
List<NumericPrediction> preds = m_predictions.get(i);
for (NumericPrediction p : preds) {
if (!Utils.isMissingValue(p.error())) {
sumSE += (p.error() * p.error());
count++;
}
}
if (m_sumOfSE[i] == 0) {
m_sumOfSE[i] = SMALL;
}
/*System.err.println("--- pred " + sumAbs + " prev " + m_sumOfAbsE[i]);
System.err.println(sumAbs / m_sumOfAbsE[i]); */
if (count == 0) {
result[i] = Utils.missingValue();
} else {
double rootMSEPrev = Math.sqrt(sumSE / count);
double rootMSE = Math.sqrt(m_sumOfSE[i] / count);
result[i] = (rootMSEPrev / rootMSE) * 100.0;
}
}
return result;
}
}