1 | #region License Information
|
---|
2 |
|
---|
3 | /* HeuristicLab
|
---|
4 | * Copyright (C) 2002-2018 Heuristic and Evolutionary Algorithms Laboratory (HEAL)
|
---|
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 |
|
---|
22 | #endregion
|
---|
23 |
|
---|
24 | using System;
|
---|
25 | using System.Collections;
|
---|
26 | using System.Collections.Generic;
|
---|
27 | using System.Linq;
|
---|
28 | using HeuristicLab.Common;
|
---|
29 | using HeuristicLab.Core;
|
---|
30 | using HeuristicLab.Data;
|
---|
31 | using HeuristicLab.Parameters;
|
---|
32 | using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
|
---|
33 | using HeuristicLab.Random;
|
---|
34 |
|
---|
35 | namespace HeuristicLab.Problems.DataAnalysis {
|
---|
36 | [StorableClass]
|
---|
37 | [Item("ClassificationSolution Impacts Calculator", "Calculation of the impacts of input variables for any classification solution")]
|
---|
38 | public sealed class ClassificationSolutionVariableImpactsCalculator : ParameterizedNamedItem {
|
---|
39 | #region Parameters/Properties
|
---|
40 | public enum ReplacementMethodEnum {
|
---|
41 | Median,
|
---|
42 | Average,
|
---|
43 | Shuffle,
|
---|
44 | Noise
|
---|
45 | }
|
---|
46 | public enum FactorReplacementMethodEnum {
|
---|
47 | Best,
|
---|
48 | Mode,
|
---|
49 | Shuffle
|
---|
50 | }
|
---|
51 | public enum DataPartitionEnum {
|
---|
52 | Training,
|
---|
53 | Test,
|
---|
54 | All
|
---|
55 | }
|
---|
56 |
|
---|
57 | private const string ReplacementParameterName = "Replacement Method";
|
---|
58 | private const string FactorReplacementParameterName = "Factor Replacement Method";
|
---|
59 | private const string DataPartitionParameterName = "DataPartition";
|
---|
60 |
|
---|
61 | public IFixedValueParameter<EnumValue<ReplacementMethodEnum>> ReplacementParameter {
|
---|
62 | get { return (IFixedValueParameter<EnumValue<ReplacementMethodEnum>>)Parameters[ReplacementParameterName]; }
|
---|
63 | }
|
---|
64 | public IFixedValueParameter<EnumValue<FactorReplacementMethodEnum>> FactorReplacementParameter {
|
---|
65 | get { return (IFixedValueParameter<EnumValue<FactorReplacementMethodEnum>>)Parameters[FactorReplacementParameterName]; }
|
---|
66 | }
|
---|
67 | public IFixedValueParameter<EnumValue<DataPartitionEnum>> DataPartitionParameter {
|
---|
68 | get { return (IFixedValueParameter<EnumValue<DataPartitionEnum>>)Parameters[DataPartitionParameterName]; }
|
---|
69 | }
|
---|
70 |
|
---|
71 | public ReplacementMethodEnum ReplacementMethod {
|
---|
72 | get { return ReplacementParameter.Value.Value; }
|
---|
73 | set { ReplacementParameter.Value.Value = value; }
|
---|
74 | }
|
---|
75 | public FactorReplacementMethodEnum FactorReplacementMethod {
|
---|
76 | get { return FactorReplacementParameter.Value.Value; }
|
---|
77 | set { FactorReplacementParameter.Value.Value = value; }
|
---|
78 | }
|
---|
79 | public DataPartitionEnum DataPartition {
|
---|
80 | get { return DataPartitionParameter.Value.Value; }
|
---|
81 | set { DataPartitionParameter.Value.Value = value; }
|
---|
82 | }
|
---|
83 | #endregion
|
---|
84 |
|
---|
85 | #region Ctor/Cloner
|
---|
86 | [StorableConstructor]
|
---|
87 | private ClassificationSolutionVariableImpactsCalculator(bool deserializing) : base(deserializing) { }
|
---|
88 | private ClassificationSolutionVariableImpactsCalculator(ClassificationSolutionVariableImpactsCalculator original, Cloner cloner)
|
---|
89 | : base(original, cloner) { }
|
---|
90 | public ClassificationSolutionVariableImpactsCalculator()
|
---|
91 | : base() {
|
---|
92 | Parameters.Add(new FixedValueParameter<EnumValue<ReplacementMethodEnum>>(ReplacementParameterName, "The replacement method for variables during impact calculation.", new EnumValue<ReplacementMethodEnum>(ReplacementMethodEnum.Shuffle)));
|
---|
93 | Parameters.Add(new FixedValueParameter<EnumValue<FactorReplacementMethodEnum>>(FactorReplacementParameterName, "The replacement method for factor variables during impact calculation.", new EnumValue<FactorReplacementMethodEnum>(FactorReplacementMethodEnum.Best)));
|
---|
94 | Parameters.Add(new FixedValueParameter<EnumValue<DataPartitionEnum>>(DataPartitionParameterName, "The data partition on which the impacts are calculated.", new EnumValue<DataPartitionEnum>(DataPartitionEnum.Training)));
|
---|
95 | }
|
---|
96 |
|
---|
97 | public override IDeepCloneable Clone(Cloner cloner) {
|
---|
98 | return new ClassificationSolutionVariableImpactsCalculator(this, cloner);
|
---|
99 | }
|
---|
100 | #endregion
|
---|
101 |
|
---|
102 | //mkommend: annoying name clash with static method, open to better naming suggestions
|
---|
103 | public IEnumerable<Tuple<string, double>> Calculate(IClassificationSolution solution) {
|
---|
104 | return CalculateImpacts(solution, ReplacementMethod, FactorReplacementMethod, DataPartition);
|
---|
105 | }
|
---|
106 |
|
---|
107 | public static IEnumerable<Tuple<string, double>> CalculateImpacts(
|
---|
108 | IClassificationSolution solution,
|
---|
109 | ReplacementMethodEnum replacementMethod = ReplacementMethodEnum.Shuffle,
|
---|
110 | FactorReplacementMethodEnum factorReplacementMethod = FactorReplacementMethodEnum.Best,
|
---|
111 | DataPartitionEnum dataPartition = DataPartitionEnum.Training) {
|
---|
112 |
|
---|
113 | IEnumerable<int> rows = GetPartitionRows(dataPartition, solution.ProblemData);
|
---|
114 | IEnumerable<double> estimatedClassValues = solution.GetEstimatedClassValues(rows);
|
---|
115 | return CalculateImpacts(solution.Model, solution.ProblemData, estimatedClassValues, rows, replacementMethod, factorReplacementMethod);
|
---|
116 | }
|
---|
117 |
|
---|
118 | public static IEnumerable<Tuple<string, double>> CalculateImpacts(
|
---|
119 | IClassificationModel model,
|
---|
120 | IClassificationProblemData problemData,
|
---|
121 | IEnumerable<double> estimatedClassValues,
|
---|
122 | IEnumerable<int> rows,
|
---|
123 | ReplacementMethodEnum replacementMethod = ReplacementMethodEnum.Shuffle,
|
---|
124 | FactorReplacementMethodEnum factorReplacementMethod = FactorReplacementMethodEnum.Best) {
|
---|
125 |
|
---|
126 | //fholzing: try and catch in case a different dataset is loaded, otherwise statement is neglectable
|
---|
127 | var missingVariables = model.VariablesUsedForPrediction.Except(problemData.Dataset.VariableNames);
|
---|
128 | if (missingVariables.Any()) {
|
---|
129 | throw new InvalidOperationException(string.Format("Can not calculate variable impacts, because the model uses inputs missing in the dataset ({0})", string.Join(", ", missingVariables)));
|
---|
130 | }
|
---|
131 | IEnumerable<double> targetValues = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, rows);
|
---|
132 | var originalQuality = CalculateQuality(targetValues, estimatedClassValues);
|
---|
133 |
|
---|
134 | var impacts = new Dictionary<string, double>();
|
---|
135 | var inputvariables = new HashSet<string>(problemData.AllowedInputVariables.Union(model.VariablesUsedForPrediction));
|
---|
136 | var modifiableDataset = ((Dataset)(problemData.Dataset).Clone()).ToModifiable();
|
---|
137 |
|
---|
138 | foreach (var inputVariable in inputvariables) {
|
---|
139 | impacts[inputVariable] = CalculateImpact(inputVariable, model, problemData, modifiableDataset, rows, replacementMethod, factorReplacementMethod, targetValues, originalQuality);
|
---|
140 | }
|
---|
141 |
|
---|
142 | return impacts.Select(i => Tuple.Create(i.Key, i.Value));
|
---|
143 | }
|
---|
144 |
|
---|
145 | public static double CalculateImpact(string variableName,
|
---|
146 | IClassificationModel model,
|
---|
147 | IClassificationProblemData problemData,
|
---|
148 | ModifiableDataset modifiableDataset,
|
---|
149 | IEnumerable<int> rows,
|
---|
150 | ReplacementMethodEnum replacementMethod = ReplacementMethodEnum.Shuffle,
|
---|
151 | FactorReplacementMethodEnum factorReplacementMethod = FactorReplacementMethodEnum.Best,
|
---|
152 | IEnumerable<double> targetValues = null,
|
---|
153 | double quality = double.NaN) {
|
---|
154 |
|
---|
155 | if (!model.VariablesUsedForPrediction.Contains(variableName)) { return 0.0; }
|
---|
156 | if (!problemData.Dataset.VariableNames.Contains(variableName)) {
|
---|
157 | throw new InvalidOperationException(string.Format("Can not calculate variable impact, because the model uses inputs missing in the dataset ({0})", variableName));
|
---|
158 | }
|
---|
159 |
|
---|
160 | if (targetValues == null) {
|
---|
161 | targetValues = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, rows);
|
---|
162 | }
|
---|
163 | if (quality == double.NaN) {
|
---|
164 | quality = CalculateQuality(model.GetEstimatedClassValues(modifiableDataset, rows), targetValues);
|
---|
165 | }
|
---|
166 |
|
---|
167 | IList originalValues = null;
|
---|
168 | IList replacementValues = GetReplacementValues(modifiableDataset, variableName, model, rows, targetValues, out originalValues, replacementMethod, factorReplacementMethod);
|
---|
169 |
|
---|
170 | double newValue = CalculateQualityForReplacement(model, modifiableDataset, variableName, originalValues, rows, replacementValues, targetValues);
|
---|
171 | double impact = quality - newValue;
|
---|
172 |
|
---|
173 | return impact;
|
---|
174 | }
|
---|
175 |
|
---|
176 | private static IList GetReplacementValues(ModifiableDataset modifiableDataset,
|
---|
177 | string variableName,
|
---|
178 | IClassificationModel model,
|
---|
179 | IEnumerable<int> rows,
|
---|
180 | IEnumerable<double> targetValues,
|
---|
181 | out IList originalValues,
|
---|
182 | ReplacementMethodEnum replacementMethod = ReplacementMethodEnum.Shuffle,
|
---|
183 | FactorReplacementMethodEnum factorReplacementMethod = FactorReplacementMethodEnum.Best) {
|
---|
184 |
|
---|
185 | IList replacementValues = null;
|
---|
186 | if (modifiableDataset.VariableHasType<double>(variableName)) {
|
---|
187 | originalValues = modifiableDataset.GetReadOnlyDoubleValues(variableName).ToList();
|
---|
188 | replacementValues = GetReplacementValuesForDouble(modifiableDataset, rows, (List<double>)originalValues, replacementMethod);
|
---|
189 | } else if (modifiableDataset.VariableHasType<string>(variableName)) {
|
---|
190 | originalValues = modifiableDataset.GetReadOnlyStringValues(variableName).ToList();
|
---|
191 | replacementValues = GetReplacementValuesForString(model, modifiableDataset, variableName, rows, (List<string>)originalValues, targetValues, factorReplacementMethod);
|
---|
192 | } else {
|
---|
193 | throw new NotSupportedException("Variable not supported");
|
---|
194 | }
|
---|
195 |
|
---|
196 | return replacementValues;
|
---|
197 | }
|
---|
198 |
|
---|
199 | private static IList GetReplacementValuesForDouble(ModifiableDataset modifiableDataset,
|
---|
200 | IEnumerable<int> rows,
|
---|
201 | List<double> originalValues,
|
---|
202 | ReplacementMethodEnum replacementMethod = ReplacementMethodEnum.Shuffle) {
|
---|
203 |
|
---|
204 | IRandom random = new FastRandom(31415);
|
---|
205 | List<double> replacementValues;
|
---|
206 | double replacementValue;
|
---|
207 |
|
---|
208 | switch (replacementMethod) {
|
---|
209 | case ReplacementMethodEnum.Median:
|
---|
210 | replacementValue = rows.Select(r => originalValues[r]).Median();
|
---|
211 | replacementValues = Enumerable.Repeat(replacementValue, modifiableDataset.Rows).ToList();
|
---|
212 | break;
|
---|
213 | case ReplacementMethodEnum.Average:
|
---|
214 | replacementValue = rows.Select(r => originalValues[r]).Average();
|
---|
215 | replacementValues = Enumerable.Repeat(replacementValue, modifiableDataset.Rows).ToList();
|
---|
216 | break;
|
---|
217 | case ReplacementMethodEnum.Shuffle:
|
---|
218 | // new var has same empirical distribution but the relation to y is broken
|
---|
219 | // prepare a complete column for the dataset
|
---|
220 | replacementValues = Enumerable.Repeat(double.NaN, modifiableDataset.Rows).ToList();
|
---|
221 | // shuffle only the selected rows
|
---|
222 | var shuffledValues = rows.Select(r => originalValues[r]).Shuffle(random).ToList();
|
---|
223 | int i = 0;
|
---|
224 | // update column values
|
---|
225 | foreach (var r in rows) {
|
---|
226 | replacementValues[r] = shuffledValues[i++];
|
---|
227 | }
|
---|
228 | break;
|
---|
229 | case ReplacementMethodEnum.Noise:
|
---|
230 | var avg = rows.Select(r => originalValues[r]).Average();
|
---|
231 | var stdDev = rows.Select(r => originalValues[r]).StandardDeviation();
|
---|
232 | // prepare a complete column for the dataset
|
---|
233 | replacementValues = Enumerable.Repeat(double.NaN, modifiableDataset.Rows).ToList();
|
---|
234 | // update column values
|
---|
235 | foreach (var r in rows) {
|
---|
236 | replacementValues[r] = NormalDistributedRandom.NextDouble(random, avg, stdDev);
|
---|
237 | }
|
---|
238 | break;
|
---|
239 |
|
---|
240 | default:
|
---|
241 | throw new ArgumentException(string.Format("ReplacementMethod {0} cannot be handled.", replacementMethod));
|
---|
242 | }
|
---|
243 |
|
---|
244 | return replacementValues;
|
---|
245 | }
|
---|
246 |
|
---|
247 | private static IList GetReplacementValuesForString(IClassificationModel model,
|
---|
248 | ModifiableDataset modifiableDataset,
|
---|
249 | string variableName,
|
---|
250 | IEnumerable<int> rows,
|
---|
251 | List<string> originalValues,
|
---|
252 | IEnumerable<double> targetValues,
|
---|
253 | FactorReplacementMethodEnum factorReplacementMethod = FactorReplacementMethodEnum.Shuffle) {
|
---|
254 |
|
---|
255 | List<string> replacementValues = null;
|
---|
256 | IRandom random = new FastRandom(31415);
|
---|
257 |
|
---|
258 | switch (factorReplacementMethod) {
|
---|
259 | case FactorReplacementMethodEnum.Best:
|
---|
260 | // try replacing with all possible values and find the best replacement value
|
---|
261 | var bestQuality = double.NegativeInfinity;
|
---|
262 | foreach (var repl in modifiableDataset.GetStringValues(variableName, rows).Distinct()) {
|
---|
263 | List<string> curReplacementValues = Enumerable.Repeat(repl, modifiableDataset.Rows).ToList();
|
---|
264 | //fholzing: this result could be used later on (theoretically), but is neglected for better readability/method consistency
|
---|
265 | var newValue = CalculateQualityForReplacement(model, modifiableDataset, variableName, originalValues, rows, curReplacementValues, targetValues);
|
---|
266 | var curQuality = newValue;
|
---|
267 |
|
---|
268 | if (curQuality > bestQuality) {
|
---|
269 | bestQuality = curQuality;
|
---|
270 | replacementValues = curReplacementValues;
|
---|
271 | }
|
---|
272 | }
|
---|
273 | break;
|
---|
274 | case FactorReplacementMethodEnum.Mode:
|
---|
275 | var mostCommonValue = rows.Select(r => originalValues[r])
|
---|
276 | .GroupBy(v => v)
|
---|
277 | .OrderByDescending(g => g.Count())
|
---|
278 | .First().Key;
|
---|
279 | replacementValues = Enumerable.Repeat(mostCommonValue, modifiableDataset.Rows).ToList();
|
---|
280 | break;
|
---|
281 | case FactorReplacementMethodEnum.Shuffle:
|
---|
282 | // new var has same empirical distribution but the relation to y is broken
|
---|
283 | // prepare a complete column for the dataset
|
---|
284 | replacementValues = Enumerable.Repeat(string.Empty, modifiableDataset.Rows).ToList();
|
---|
285 | // shuffle only the selected rows
|
---|
286 | var shuffledValues = rows.Select(r => originalValues[r]).Shuffle(random).ToList();
|
---|
287 | int i = 0;
|
---|
288 | // update column values
|
---|
289 | foreach (var r in rows) {
|
---|
290 | replacementValues[r] = shuffledValues[i++];
|
---|
291 | }
|
---|
292 | break;
|
---|
293 | default:
|
---|
294 | throw new ArgumentException(string.Format("FactorReplacementMethod {0} cannot be handled.", factorReplacementMethod));
|
---|
295 | }
|
---|
296 |
|
---|
297 | return replacementValues;
|
---|
298 | }
|
---|
299 |
|
---|
300 | private static double CalculateQualityForReplacement(
|
---|
301 | IClassificationModel model,
|
---|
302 | ModifiableDataset modifiableDataset,
|
---|
303 | string variableName,
|
---|
304 | IList originalValues,
|
---|
305 | IEnumerable<int> rows,
|
---|
306 | IList replacementValues,
|
---|
307 | IEnumerable<double> targetValues) {
|
---|
308 |
|
---|
309 | modifiableDataset.ReplaceVariable(variableName, replacementValues);
|
---|
310 | var discModel = model as IDiscriminantFunctionClassificationModel;
|
---|
311 | if (discModel != null) {
|
---|
312 | var problemData = new ClassificationProblemData(modifiableDataset, modifiableDataset.VariableNames, model.TargetVariable);
|
---|
313 | discModel.RecalculateModelParameters(problemData, rows);
|
---|
314 | }
|
---|
315 |
|
---|
316 | //mkommend: ToList is used on purpose to avoid lazy evaluation that could result in wrong estimates due to variable replacements
|
---|
317 | var estimates = model.GetEstimatedClassValues(modifiableDataset, rows).ToList();
|
---|
318 | var ret = CalculateQuality(targetValues, estimates);
|
---|
319 | modifiableDataset.ReplaceVariable(variableName, originalValues);
|
---|
320 |
|
---|
321 | return ret;
|
---|
322 | }
|
---|
323 |
|
---|
324 | public static double CalculateQuality(IEnumerable<double> targetValues, IEnumerable<double> estimatedClassValues) {
|
---|
325 | OnlineCalculatorError errorState;
|
---|
326 | var ret = OnlineAccuracyCalculator.Calculate(targetValues, estimatedClassValues, out errorState);
|
---|
327 | if (errorState != OnlineCalculatorError.None) { throw new InvalidOperationException("Error during calculation with replaced inputs."); }
|
---|
328 | return ret;
|
---|
329 | }
|
---|
330 |
|
---|
331 | public static IEnumerable<int> GetPartitionRows(DataPartitionEnum dataPartition, IClassificationProblemData problemData) {
|
---|
332 | IEnumerable<int> rows;
|
---|
333 |
|
---|
334 | switch (dataPartition) {
|
---|
335 | case DataPartitionEnum.All:
|
---|
336 | rows = problemData.AllIndices;
|
---|
337 | break;
|
---|
338 | case DataPartitionEnum.Test:
|
---|
339 | rows = problemData.TestIndices;
|
---|
340 | break;
|
---|
341 | case DataPartitionEnum.Training:
|
---|
342 | rows = problemData.TrainingIndices;
|
---|
343 | break;
|
---|
344 | default:
|
---|
345 | throw new NotSupportedException("DataPartition not supported");
|
---|
346 | }
|
---|
347 |
|
---|
348 | return rows;
|
---|
349 | }
|
---|
350 | }
|
---|
351 | } |
---|