source: trunk/sources/HeuristicLab.Problems.DataAnalysis/3.4/Implementation/Classification/ThresholdCalculators/AccuracyMaximizationThresholdCalculator.cs @ 8636

Last change on this file since 8636 was 8636, checked in by mkommend, 7 years ago

#1924:

  • Changed the accuracy threshold calculator to eliminate the necessity the the class values are ordered.
  • Adapted the symbolic classification simplifier to work with all ISymbolicClassificationModels.
  • Corrected ROCCurvesView to also work if the class values are not sorted.
File size: 7.1 KB
Line 
1#region License Information
2/* HeuristicLab
3 * Copyright (C) 2002-2012 Heuristic and Evolutionary Algorithms Laboratory (HEAL)
4 *
5 * This file is part of HeuristicLab.
6 *
7 * HeuristicLab is free software: you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation, either version 3 of the License, or
10 * (at your option) any later version.
11 *
12 * HeuristicLab is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15 * GNU General Public License for more details.
16 *
17 * You should have received a copy of the GNU General Public License
18 * along with HeuristicLab. If not, see <http://www.gnu.org/licenses/>.
19 */
20#endregion
21
22using System;
23using System.Collections.Generic;
24using System.Linq;
25using HeuristicLab.Common;
26using HeuristicLab.Core;
27using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
28
29namespace HeuristicLab.Problems.DataAnalysis {
30  /// <summary>
31  /// Represents a threshold calculator that maximizes the weighted accuracy of the classifcation model.
32  /// </summary>
33  [StorableClass]
34  [Item("AccuracyMaximizationThresholdCalculator", "Represents a threshold calculator that maximizes the weighted accuracy of the classifcation model.")]
35  public class AccuracyMaximizationThresholdCalculator : ThresholdCalculator {
36
37    [StorableConstructor]
38    protected AccuracyMaximizationThresholdCalculator(bool deserializing) : base(deserializing) { }
39    protected AccuracyMaximizationThresholdCalculator(AccuracyMaximizationThresholdCalculator original, Cloner cloner)
40      : base(original, cloner) {
41    }
42    public AccuracyMaximizationThresholdCalculator()
43      : base() {
44    }
45
46    public override IDeepCloneable Clone(Cloner cloner) {
47      return new AccuracyMaximizationThresholdCalculator(this, cloner);
48    }
49
50    public override void Calculate(IClassificationProblemData problemData, IEnumerable<double> estimatedValues, IEnumerable<double> targetClassValues, out double[] classValues, out double[] thresholds) {
51      AccuracyMaximizationThresholdCalculator.CalculateThresholds(problemData, estimatedValues, targetClassValues, out classValues, out thresholds);
52    }
53
54    public static void CalculateThresholds(IClassificationProblemData problemData, IEnumerable<double> estimatedValues, IEnumerable<double> targetClassValues, out double[] classValues, out double[] thresholds) {
55      const int slices = 100;
56      const double minThresholdInc = 10e-5; // necessary to prevent infinite loop when maxEstimated - minEstimated is effectively zero (constant model)
57      List<double> estimatedValuesList = estimatedValues.ToList();
58      double maxEstimatedValue = estimatedValuesList.Max();
59      double minEstimatedValue = estimatedValuesList.Min();
60      double thresholdIncrement = Math.Max((maxEstimatedValue - minEstimatedValue) / slices, minThresholdInc);
61      var estimatedAndTargetValuePairs =
62        estimatedValuesList.Zip(targetClassValues, (x, y) => new { EstimatedValue = x, TargetClassValue = y })
63        .OrderBy(x => x.EstimatedValue).ToList();
64
65      classValues = estimatedAndTargetValuePairs.GroupBy(x => x.TargetClassValue)
66        .Select(x => new { Median = x.Select(y => y.EstimatedValue).Median(), Class = x.Key })
67        .OrderBy(x => x.Median).Select(x => x.Class).ToArray();
68
69      int nClasses = classValues.Length;
70      thresholds = new double[nClasses];
71      thresholds[0] = double.NegativeInfinity;
72
73      // incrementally calculate accuracy of all possible thresholds
74      for (int i = 1; i < thresholds.Length; i++) {
75        double lowerThreshold = thresholds[i - 1];
76        double actualThreshold = Math.Max(lowerThreshold, minEstimatedValue);
77        double lowestBestThreshold = double.NaN;
78        double highestBestThreshold = double.NaN;
79        double bestClassificationScore = double.PositiveInfinity;
80        bool seriesOfEqualClassificationScores = false;
81
82        while (actualThreshold < maxEstimatedValue) {
83          double classificationScore = 0.0;
84
85          foreach (var pair in estimatedAndTargetValuePairs) {
86            //all positives
87            if (pair.TargetClassValue.IsAlmost(classValues[i - 1])) {
88              if (pair.EstimatedValue > lowerThreshold && pair.EstimatedValue <= actualThreshold)
89                //true positive
90                classificationScore += problemData.GetClassificationPenalty(pair.TargetClassValue, pair.TargetClassValue);
91              else
92                //false negative
93                classificationScore += problemData.GetClassificationPenalty(pair.TargetClassValue, classValues[i]);
94            }
95              //all negatives
96            else {
97              //false positive
98              if (pair.EstimatedValue > lowerThreshold && pair.EstimatedValue <= actualThreshold)
99                classificationScore += problemData.GetClassificationPenalty(pair.TargetClassValue, classValues[i - 1]);
100              else if (pair.EstimatedValue <= lowerThreshold)
101                classificationScore += problemData.GetClassificationPenalty(pair.TargetClassValue, classValues[i - 2]);
102              else if (pair.EstimatedValue > actualThreshold) {
103                if (pair.TargetClassValue < classValues[i - 1]) //negative in wrong class, consider upper class
104                  classificationScore += problemData.GetClassificationPenalty(pair.TargetClassValue, classValues[i]);
105                else //true negative, must be optimized by the other thresholds
106                  classificationScore += problemData.GetClassificationPenalty(pair.TargetClassValue, pair.TargetClassValue);
107              }
108            }
109          }
110
111          //new best classification score found
112          if (classificationScore < bestClassificationScore) {
113            bestClassificationScore = classificationScore;
114            lowestBestThreshold = actualThreshold;
115            highestBestThreshold = actualThreshold;
116            seriesOfEqualClassificationScores = true;
117          }
118            //equal classification scores => if seriesOfEqualClassifcationScores == true update highest threshold
119          else if (Math.Abs(classificationScore - bestClassificationScore) < double.Epsilon && seriesOfEqualClassificationScores)
120            highestBestThreshold = actualThreshold;
121          //worse classificatoin score found reset seriesOfEqualClassifcationScores
122          else seriesOfEqualClassificationScores = false;
123
124          actualThreshold += thresholdIncrement;
125        }
126        //scale lowest thresholds and highest found optimal threshold according to the misclassification matrix
127        double falseNegativePenalty = problemData.GetClassificationPenalty(classValues[i], classValues[i - 1]);
128        double falsePositivePenalty = problemData.GetClassificationPenalty(classValues[i - 1], classValues[i]);
129        thresholds[i] = (lowestBestThreshold * falsePositivePenalty + highestBestThreshold * falseNegativePenalty) / (falseNegativePenalty + falsePositivePenalty);
130      }
131    }
132  }
133}
Note: See TracBrowser for help on using the repository browser.