Free cookie consent management tool by TermsFeed Policy Generator

source: stable/HeuristicLab.Algorithms.DataAnalysis/3.4/GradientBoostedTrees/GradientBoostedTreesAlgorithm.cs @ 17097

Last change on this file since 17097 was 17097, checked in by mkommend, 5 years ago

#2520: Merged 16565 - 16579 into stable.

File size: 13.4 KB
Line 
1#region License Information
2/* HeuristicLab
3 * Copyright (C) 2002-2019 Heuristic and Evolutionary Algorithms Laboratory (HEAL)
4 * and the BEACON Center for the Study of Evolution in Action.
5 *
6 * This file is part of HeuristicLab.
7 *
8 * HeuristicLab is free software: you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation, either version 3 of the License, or
11 * (at your option) any later version.
12 *
13 * HeuristicLab is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16 * GNU General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License
19 * along with HeuristicLab. If not, see <http://www.gnu.org/licenses/>.
20 */
21#endregion
22
23using System;
24using System.Linq;
25using System.Threading;
26using HeuristicLab.Analysis;
27using HeuristicLab.Common;
28using HeuristicLab.Core;
29using HeuristicLab.Data;
30using HeuristicLab.Optimization;
31using HeuristicLab.Parameters;
32using HEAL.Attic;
33using HeuristicLab.PluginInfrastructure;
34using HeuristicLab.Problems.DataAnalysis;
35
36namespace HeuristicLab.Algorithms.DataAnalysis {
37  [Item("Gradient Boosted Trees (GBT)", "Gradient boosted trees algorithm. Specific implementation of gradient boosting for regression trees. Friedman, J. \"Greedy Function Approximation: A Gradient Boosting Machine\", IMS 1999 Reitz Lecture.")]
38  [StorableType("8CCB55BD-4935-4868-855F-D3E5D55127AA")]
39  [Creatable(CreatableAttribute.Categories.DataAnalysisRegression, Priority = 125)]
40  public class GradientBoostedTreesAlgorithm : FixedDataAnalysisAlgorithm<IRegressionProblem> {
41    #region ParameterNames
42    private const string IterationsParameterName = "Iterations";
43    private const string MaxSizeParameterName = "Maximum Tree Size";
44    private const string NuParameterName = "Nu";
45    private const string RParameterName = "R";
46    private const string MParameterName = "M";
47    private const string SeedParameterName = "Seed";
48    private const string SetSeedRandomlyParameterName = "SetSeedRandomly";
49    private const string LossFunctionParameterName = "LossFunction";
50    private const string UpdateIntervalParameterName = "UpdateInterval";
51    private const string CreateSolutionParameterName = "CreateSolution";
52    #endregion
53
54    #region ParameterProperties
55    public IFixedValueParameter<IntValue> IterationsParameter {
56      get { return (IFixedValueParameter<IntValue>)Parameters[IterationsParameterName]; }
57    }
58    public IFixedValueParameter<IntValue> MaxSizeParameter {
59      get { return (IFixedValueParameter<IntValue>)Parameters[MaxSizeParameterName]; }
60    }
61    public IFixedValueParameter<DoubleValue> NuParameter {
62      get { return (IFixedValueParameter<DoubleValue>)Parameters[NuParameterName]; }
63    }
64    public IFixedValueParameter<DoubleValue> RParameter {
65      get { return (IFixedValueParameter<DoubleValue>)Parameters[RParameterName]; }
66    }
67    public IFixedValueParameter<DoubleValue> MParameter {
68      get { return (IFixedValueParameter<DoubleValue>)Parameters[MParameterName]; }
69    }
70    public IFixedValueParameter<IntValue> SeedParameter {
71      get { return (IFixedValueParameter<IntValue>)Parameters[SeedParameterName]; }
72    }
73    public FixedValueParameter<BoolValue> SetSeedRandomlyParameter {
74      get { return (FixedValueParameter<BoolValue>)Parameters[SetSeedRandomlyParameterName]; }
75    }
76    public IConstrainedValueParameter<ILossFunction> LossFunctionParameter {
77      get { return (IConstrainedValueParameter<ILossFunction>)Parameters[LossFunctionParameterName]; }
78    }
79    public IFixedValueParameter<IntValue> UpdateIntervalParameter {
80      get { return (IFixedValueParameter<IntValue>)Parameters[UpdateIntervalParameterName]; }
81    }
82    public IFixedValueParameter<BoolValue> CreateSolutionParameter {
83      get { return (IFixedValueParameter<BoolValue>)Parameters[CreateSolutionParameterName]; }
84    }
85    #endregion
86
87    #region Properties
88    public int Iterations {
89      get { return IterationsParameter.Value.Value; }
90      set { IterationsParameter.Value.Value = value; }
91    }
92    public int Seed {
93      get { return SeedParameter.Value.Value; }
94      set { SeedParameter.Value.Value = value; }
95    }
96    public bool SetSeedRandomly {
97      get { return SetSeedRandomlyParameter.Value.Value; }
98      set { SetSeedRandomlyParameter.Value.Value = value; }
99    }
100    public int MaxSize {
101      get { return MaxSizeParameter.Value.Value; }
102      set { MaxSizeParameter.Value.Value = value; }
103    }
104    public double Nu {
105      get { return NuParameter.Value.Value; }
106      set { NuParameter.Value.Value = value; }
107    }
108    public double R {
109      get { return RParameter.Value.Value; }
110      set { RParameter.Value.Value = value; }
111    }
112    public double M {
113      get { return MParameter.Value.Value; }
114      set { MParameter.Value.Value = value; }
115    }
116    public bool CreateSolution {
117      get { return CreateSolutionParameter.Value.Value; }
118      set { CreateSolutionParameter.Value.Value = value; }
119    }
120    #endregion
121
122    #region ResultsProperties
123    private double ResultsBestQuality {
124      get { return ((DoubleValue)Results["Best Quality"].Value).Value; }
125      set { ((DoubleValue)Results["Best Quality"].Value).Value = value; }
126    }
127    private DataTable ResultsQualities {
128      get { return ((DataTable)Results["Qualities"].Value); }
129    }
130    #endregion
131
132    [StorableConstructor]
133    protected GradientBoostedTreesAlgorithm(StorableConstructorFlag _) : base(_) { }
134
135    protected GradientBoostedTreesAlgorithm(GradientBoostedTreesAlgorithm original, Cloner cloner)
136      : base(original, cloner) {
137    }
138
139    public override IDeepCloneable Clone(Cloner cloner) {
140      return new GradientBoostedTreesAlgorithm(this, cloner);
141    }
142
143    public GradientBoostedTreesAlgorithm() {
144      Problem = new RegressionProblem(); // default problem
145
146      Parameters.Add(new FixedValueParameter<IntValue>(IterationsParameterName, "Number of iterations (set as high as possible, adjust in combination with nu, when increasing iterations also decrease nu)", new IntValue(1000)));
147      Parameters.Add(new FixedValueParameter<IntValue>(SeedParameterName, "The random seed used to initialize the new pseudo random number generator.", new IntValue(0)));
148      Parameters.Add(new FixedValueParameter<BoolValue>(SetSeedRandomlyParameterName, "True if the random seed should be set to a random value, otherwise false.", new BoolValue(true)));
149      Parameters.Add(new FixedValueParameter<IntValue>(MaxSizeParameterName, "Maximal size of the tree learned in each step (prefer smaller sizes if possible)", new IntValue(10)));
150      Parameters.Add(new FixedValueParameter<DoubleValue>(RParameterName, "Ratio of training rows selected randomly in each step (0 < R <= 1)", new DoubleValue(0.5)));
151      Parameters.Add(new FixedValueParameter<DoubleValue>(MParameterName, "Ratio of variables selected randomly in each step (0 < M <= 1)", new DoubleValue(0.5)));
152      Parameters.Add(new FixedValueParameter<DoubleValue>(NuParameterName, "Learning rate nu (step size for the gradient update, should be small 0 < nu < 0.1)", new DoubleValue(0.002)));
153      Parameters.Add(new FixedValueParameter<IntValue>(UpdateIntervalParameterName, "", new IntValue(100)));
154      Parameters[UpdateIntervalParameterName].Hidden = true;
155      Parameters.Add(new FixedValueParameter<BoolValue>(CreateSolutionParameterName, "Flag that indicates if a solution should be produced at the end of the run", new BoolValue(true)));
156      Parameters[CreateSolutionParameterName].Hidden = true;
157
158      var lossFunctions = ApplicationManager.Manager.GetInstances<ILossFunction>();
159      Parameters.Add(new ConstrainedValueParameter<ILossFunction>(LossFunctionParameterName, "The loss function", new ItemSet<ILossFunction>(lossFunctions)));
160      LossFunctionParameter.Value = LossFunctionParameter.ValidValues.First(f => f.ToString().Contains("Squared")); // squared error loss is the default
161    }
162
163    [StorableHook(HookType.AfterDeserialization)]
164    private void AfterDeserialization() {
165      // BackwardsCompatibility3.4
166      #region Backwards compatible code, remove with 3.5
167      // parameter type has been changed
168      var lossFunctionParam = Parameters[LossFunctionParameterName] as ConstrainedValueParameter<StringValue>;
169      if (lossFunctionParam != null) {
170        Parameters.Remove(LossFunctionParameterName);
171        var selectedValue = lossFunctionParam.Value; // to be restored below
172
173        var lossFunctions = ApplicationManager.Manager.GetInstances<ILossFunction>();
174        Parameters.Add(new ConstrainedValueParameter<ILossFunction>(LossFunctionParameterName, "The loss function", new ItemSet<ILossFunction>(lossFunctions)));
175        // try to restore selected value
176        var selectedLossFunction =
177          LossFunctionParameter.ValidValues.FirstOrDefault(f => f.ToString() == selectedValue.Value);
178        if (selectedLossFunction != null) {
179          LossFunctionParameter.Value = selectedLossFunction;
180        } else {
181          LossFunctionParameter.Value = LossFunctionParameter.ValidValues.First(f => f.ToString().Contains("Squared")); // default: SE
182        }
183      }
184      #endregion
185    }
186
187    protected override void Run(CancellationToken cancellationToken) {
188      // Set up the algorithm
189      if (SetSeedRandomly) Seed = Random.RandomSeedGenerator.GetSeed();
190
191      // Set up the results display
192      var iterations = new IntValue(0);
193      Results.Add(new Result("Iterations", iterations));
194
195      var table = new DataTable("Qualities");
196      table.Rows.Add(new DataRow("Loss (train)"));
197      table.Rows.Add(new DataRow("Loss (test)"));
198      table.Rows["Loss (train)"].VisualProperties.StartIndexZero = true;
199      table.Rows["Loss (test)"].VisualProperties.StartIndexZero = true;
200
201      Results.Add(new Result("Qualities", table));
202      var curLoss = new DoubleValue();
203      Results.Add(new Result("Loss (train)", curLoss));
204
205      // init
206      var problemData = (IRegressionProblemData)Problem.ProblemData.Clone();
207      var lossFunction = LossFunctionParameter.Value;
208      var state = GradientBoostedTreesAlgorithmStatic.CreateGbmState(problemData, lossFunction, (uint)Seed, MaxSize, R, M, Nu);
209
210      var updateInterval = UpdateIntervalParameter.Value.Value;
211      // Loop until iteration limit reached or canceled.
212      for (int i = 0; i < Iterations; i++) {
213        cancellationToken.ThrowIfCancellationRequested();
214
215        GradientBoostedTreesAlgorithmStatic.MakeStep(state);
216
217        // iteration results
218        if (i % updateInterval == 0) {
219          curLoss.Value = state.GetTrainLoss();
220          table.Rows["Loss (train)"].Values.Add(curLoss.Value);
221          table.Rows["Loss (test)"].Values.Add(state.GetTestLoss());
222          iterations.Value = i;
223        }
224      }
225
226      // final results
227      iterations.Value = Iterations;
228      curLoss.Value = state.GetTrainLoss();
229      table.Rows["Loss (train)"].Values.Add(curLoss.Value);
230      table.Rows["Loss (test)"].Values.Add(state.GetTestLoss());
231
232      // produce variable relevance
233      var orderedImpacts = state.GetVariableRelevance().Select(t => new { name = t.Key, impact = t.Value }).ToList();
234
235      var impacts = new DoubleMatrix();
236      var matrix = impacts as IStringConvertibleMatrix;
237      matrix.Rows = orderedImpacts.Count;
238      matrix.RowNames = orderedImpacts.Select(x => x.name);
239      matrix.Columns = 1;
240      matrix.ColumnNames = new string[] { "Relative variable relevance" };
241
242      int rowIdx = 0;
243      foreach (var p in orderedImpacts) {
244        matrix.SetValue(string.Format("{0:N2}", p.impact), rowIdx++, 0);
245      }
246
247      Results.Add(new Result("Variable relevance", impacts));
248      Results.Add(new Result("Loss (test)", new DoubleValue(state.GetTestLoss())));
249
250      // produce solution
251      if (CreateSolution) {
252        var model = state.GetModel();
253
254        // for logistic regression we produce a classification solution
255        if (lossFunction is LogisticRegressionLoss) {
256          var classificationModel = new DiscriminantFunctionClassificationModel(model,
257            new AccuracyMaximizationThresholdCalculator());
258          var classificationProblemData = new ClassificationProblemData(problemData.Dataset,
259            problemData.AllowedInputVariables, problemData.TargetVariable, problemData.Transformations);
260          classificationProblemData.TrainingPartition.Start = Problem.ProblemData.TrainingPartition.Start;
261          classificationProblemData.TrainingPartition.End = Problem.ProblemData.TrainingPartition.End;
262          classificationProblemData.TestPartition.Start = Problem.ProblemData.TestPartition.Start;
263          classificationProblemData.TestPartition.End = Problem.ProblemData.TestPartition.End;
264
265          classificationModel.SetThresholdsAndClassValues(new double[] { double.NegativeInfinity, 0.0 }, new[] { 0.0, 1.0 });
266
267
268          var classificationSolution = new DiscriminantFunctionClassificationSolution(classificationModel, classificationProblemData);
269          Results.Add(new Result("Solution", classificationSolution));
270        } else {
271          // otherwise we produce a regression solution
272          Results.Add(new Result("Solution", new GradientBoostedTreesSolution(model, problemData)));
273        }
274      }
275    }
276  }
277}
Note: See TracBrowser for help on using the repository browser.