/// <summary>
/// Computes all per-label AUCs as well as the micro- and macro-averaged AUCs.
/// </summary>
/// <param name="confusionMatrix">The confusion matrix.</param>
/// <param name="evaluator">The classifier evaluator.</param>
/// <param name="x">The x vector of the ground truth.</param>
/// <param name="y">The y of the ground truth.</param>
/// <param name="yPredicDistrib">The predictive distributions.</param>
/// <param name="microAuc">The micro-averaged area under the receiver operating characteristic curve.</param>
/// <param name="macroAuc">The macro-averaged area under the receiver operating characteristic curve.</param>
/// <param name="macroAucClassLabelCount">The number of class labels for which the AUC if defined.</param>
/// <returns>The area under the receiver operating characteristic curve for each class label.</returns>
/// <remarks>Adapted from MicrosoftResearch.Infer.Learners</remarks>
private IDictionary <string, double> ComputeLabelAuc(
ConfusionMatrix <string> confusionMatrix,
ClassifierEvaluator <IList <Vector>, int, IList <string>, string> evaluator,
Vector[] x,
IList <string> y,
IEnumerable <IDictionary <string, double> > yPredicDistrib,
out double microAuc,
out double macroAuc,
out int macroAucClassLabelCount)
{
int instanceCount = yPredicDistrib.Count();
var classLabels = confusionMatrix.ClassLabelSet.Elements.ToArray();
int classLabelCount = classLabels.Length;
var labelAuc = new Dictionary <string, double>();
// Compute per-label AUC
macroAucClassLabelCount = classLabelCount;
foreach (var classLabel in classLabels)
{
// One versus rest
double auc;
try
{
auc = evaluator.AreaUnderRocCurve(classLabel, x, y, yPredicDistrib);
}
catch (ArgumentException)
{
auc = double.NaN;
macroAucClassLabelCount--;
}
labelAuc.Add(classLabel, auc);
}
// Compute micro- and macro-averaged AUC
microAuc = 0;
macroAuc = 0;
foreach (var label in classLabels)
{
if (double.IsNaN(labelAuc[label]))
{
continue;
}
microAuc += confusionMatrix.TrueLabelCount(label) * labelAuc[label] / instanceCount;
macroAuc += labelAuc[label] / macroAucClassLabelCount;
}
return(labelAuc);
}
/// <summary>
/// Computes all per-label AUCs as well as the micro- and macro-averaged AUCs.
/// </summary>
/// <param name="confusionMatrix">The confusion matrix.</param>
/// <param name="evaluator">The classifier evaluator.</param>
/// <param name="groundTruth">The ground truth.</param>
/// <param name="predictiveDistributions">The predictive distributions.</param>
/// <param name="microAuc">The micro-averaged area under the receiver operating characteristic curve.</param>
/// <param name="macroAuc">The macro-averaged area under the receiver operating characteristic curve.</param>
/// <param name="macroAucClassLabelCount">The number of class labels for which the AUC if defined.</param>
/// <returns>The area under the receiver operating characteristic curve for each class label.</returns>
private IDictionary <string, double> ComputeLabelAuc(
ConfusionMatrix <string> confusionMatrix,
ClassifierEvaluator <IList <LabeledFeatureValues>, LabeledFeatureValues, IList <LabelDistribution>, string> evaluator,
IList <LabeledFeatureValues> groundTruth,
ICollection <IDictionary <string, double> > predictiveDistributions,
out double microAuc,
out double macroAuc,
out int macroAucClassLabelCount)
{
int instanceCount = predictiveDistributions.Count;
var classLabels = confusionMatrix.ClassLabelSet.Elements.ToArray();
int classLabelCount = classLabels.Length;
var labelAuc = new Dictionary <string, double>();
// Compute per-label AUC
macroAucClassLabelCount = classLabelCount;
foreach (var classLabel in classLabels)
{
// One versus rest
double auc;
try
{
auc = evaluator.AreaUnderRocCurve(classLabel, groundTruth, predictiveDistributions);
}
catch (ArgumentException)
{
auc = double.NaN;
macroAucClassLabelCount--;
}
labelAuc.Add(classLabel, auc);
}
// Compute micro- and macro-averaged AUC
microAuc = 0;
macroAuc = 0;
foreach (var label in classLabels)
{
if (double.IsNaN(labelAuc[label]))
{
continue;
}
microAuc += confusionMatrix.TrueLabelCount(label) * labelAuc[label] / instanceCount;
macroAuc += labelAuc[label] / macroAucClassLabelCount;
}
return(labelAuc);
}
/// <summary>
/// Writes the evaluation results to a file with the specified name.
/// </summary>
/// <param name="writer">The name of the file to write the report to.</param>
/// <param name="evaluator">The classifier evaluator.</param>
/// <param name="x">The x vector of the ground truth.</param>
/// <param name="y">The y of the ground truth.</param>
/// <param name="yPredicDistrib">The predictive distributions.</param>
/// <param name="yPredicLabel">The predicted labels.</param>
/// <remarks>Adapted from MicrosoftResearch.Infer.Learners</remarks>
private void WriteReport(
StreamWriter writer,
ClassifierEvaluator <IList <Vector>, int, IList <string>, string> evaluator,
Vector[] x,
IList <string> y,
IEnumerable <IDictionary <string, double> > yPredicDistrib,
IEnumerable <string> yPredicLabel)
{
// Compute confusion matrix
var confusionMatrix = evaluator.ConfusionMatrix(x, y, yPredicLabel);
// Compute mean negative log probability
double meanNegativeLogProbability =
evaluator.Evaluate(x, y, yPredicDistrib, Metrics.NegativeLogProbability) / yPredicDistrib.Count();
// Compute M-measure (averaged pairwise AUC)
IDictionary <string, IDictionary <string, double> > aucMatrix;
double auc = evaluator.AreaUnderRocCurve(x, y, yPredicDistrib, out aucMatrix);
// Compute per-label AUC as well as micro- and macro-averaged AUC
double microAuc;
double macroAuc;
int macroAucClassLabelCount;
var labelAuc = this.ComputeLabelAuc(
confusionMatrix,
evaluator,
x,
y,
yPredicDistrib,
out microAuc,
out macroAuc,
out macroAucClassLabelCount);
// Instance-averaged performance
this.WriteInstanceAveragedPerformance(writer, confusionMatrix, meanNegativeLogProbability, microAuc);
// Class-averaged performance
this.WriteClassAveragedPerformance(writer, confusionMatrix, auc, macroAuc, macroAucClassLabelCount);
// Performance on individual classes
this.WriteIndividualClassPerformance(writer, confusionMatrix, labelAuc);
// Confusion matrix
this.WriteConfusionMatrix(writer, confusionMatrix);
// Pairwise AUC
this.WriteAucMatrix(writer, aucMatrix);
}
/// <summary>
/// Writes the evaluation results to a file with the specified name.
/// </summary>
/// <param name="writer">The name of the file to write the report to.</param>
/// <param name="evaluator">The classifier evaluator.</param>
/// <param name="groundTruth">The ground truth.</param>
/// <param name="predictiveDistributions">The predictive distributions.</param>
/// <param name="predictedLabels">The predicted labels.</param>
private void WriteReport(
StreamWriter writer,
ClassifierEvaluator <IList <LabeledFeatureValues>, LabeledFeatureValues, IList <LabelDistribution>, string> evaluator,
IList <LabeledFeatureValues> groundTruth,
ICollection <IDictionary <string, double> > predictiveDistributions,
IEnumerable <string> predictedLabels)
{
// Compute confusion matrix
var confusionMatrix = evaluator.ConfusionMatrix(groundTruth, predictedLabels);
// Compute mean negative log probability
double meanNegativeLogProbability =
evaluator.Evaluate(groundTruth, predictiveDistributions, Metrics.NegativeLogProbability) / predictiveDistributions.Count;
// Compute M-measure (averaged pairwise AUC)
IDictionary <string, IDictionary <string, double> > aucMatrix;
double auc = evaluator.AreaUnderRocCurve(groundTruth, predictiveDistributions, out aucMatrix);
// Compute per-label AUC as well as micro- and macro-averaged AUC
double microAuc;
double macroAuc;
int macroAucClassLabelCount;
var labelAuc = this.ComputeLabelAuc(
confusionMatrix,
evaluator,
groundTruth,
predictiveDistributions,
out microAuc,
out macroAuc,
out macroAucClassLabelCount);
// Instance-averaged performance
this.WriteInstanceAveragedPerformance(writer, confusionMatrix, meanNegativeLogProbability, microAuc);
// Class-averaged performance
this.WriteClassAveragedPerformance(writer, confusionMatrix, auc, macroAuc, macroAucClassLabelCount);
// Performance on individual classes
this.WriteIndividualClassPerformance(writer, confusionMatrix, labelAuc);
// Confusion matrix
this.WriteConfusionMatrix(writer, confusionMatrix);
// Pairwise AUC
this.WriteAucMatrix(writer, aucMatrix);
}
/// <summary>
/// Runs the module.
/// </summary>
/// <param name="args">The command line arguments for the module.</param>
/// <param name="usagePrefix">The prefix to print before the usage string.</param>
/// <returns>True if the run was successful, false otherwise.</returns>
public override bool Run(string[] args, string usagePrefix)
{
string dataSetFile = string.Empty;
string resultsFile = string.Empty;
int crossValidationFoldCount = 5;
int iterationCount = BayesPointMachineClassifierTrainingSettings.IterationCountDefault;
int batchCount = BayesPointMachineClassifierTrainingSettings.BatchCountDefault;
bool computeModelEvidence = BayesPointMachineClassifierTrainingSettings.ComputeModelEvidenceDefault;
var parser = new CommandLineParser();
parser.RegisterParameterHandler("--data-set", "FILE", "File with training data", v => dataSetFile = v, CommandLineParameterType.Required);
parser.RegisterParameterHandler("--results", "FILE", "File with cross-validation results", v => resultsFile = v, CommandLineParameterType.Required);
parser.RegisterParameterHandler("--folds", "NUM", "Number of cross-validation folds (defaults to " + crossValidationFoldCount + ")", v => crossValidationFoldCount = v, CommandLineParameterType.Optional);
parser.RegisterParameterHandler("--iterations", "NUM", "Number of training algorithm iterations (defaults to " + iterationCount + ")", v => iterationCount = v, CommandLineParameterType.Optional);
parser.RegisterParameterHandler("--batches", "NUM", "Number of batches to split the training data into (defaults to " + batchCount + ")", v => batchCount = v, CommandLineParameterType.Optional);
parser.RegisterParameterHandler("--compute-evidence", "Compute model evidence (defaults to " + computeModelEvidence + ")", () => computeModelEvidence = true);
if (!parser.TryParse(args, usagePrefix))
{
return(false);
}
// Load and shuffle data
var dataSet = ClassifierPersistenceUtils.LoadLabeledFeatureValues(dataSetFile);
BayesPointMachineClassifierModuleUtilities.WriteDataSetInfo(dataSet);
Rand.Restart(562);
Rand.Shuffle(dataSet);
// Create evaluator
var evaluatorMapping = Mappings.Classifier.ForEvaluation();
var evaluator = new ClassifierEvaluator <IList <LabeledFeatureValues>, LabeledFeatureValues, IList <LabelDistribution>, string>(evaluatorMapping);
// Create performance metrics
var accuracy = new List <double>();
var negativeLogProbability = new List <double>();
var auc = new List <double>();
var evidence = new List <double>();
var iterationCounts = new List <double>();
var trainingTime = new List <double>();
// Run cross-validation
int validationSetSize = dataSet.Count / crossValidationFoldCount;
Console.WriteLine("Running {0}-fold cross-validation on {1}", crossValidationFoldCount, dataSetFile);
// TODO: Use chained mapping to implement cross-validation
for (int fold = 0; fold < crossValidationFoldCount; fold++)
{
// Construct training and validation sets for fold
int validationSetStart = fold * validationSetSize;
int validationSetEnd = (fold + 1 == crossValidationFoldCount)
? dataSet.Count
: (fold + 1) * validationSetSize;
var trainingSet = new List <LabeledFeatureValues>();
var validationSet = new List <LabeledFeatureValues>();
for (int instance = 0; instance < dataSet.Count; instance++)
{
if (validationSetStart <= instance && instance < validationSetEnd)
{
validationSet.Add(dataSet[instance]);
}
else
{
trainingSet.Add(dataSet[instance]);
}
}
// Print info
Console.WriteLine(" Fold {0} [validation set instances {1} - {2}]", fold + 1, validationSetStart, validationSetEnd - 1);
// Create classifier
var classifier = BayesPointMachineClassifier.CreateBinaryClassifier(Mappings.Classifier);
classifier.Settings.Training.IterationCount = iterationCount;
classifier.Settings.Training.BatchCount = batchCount;
classifier.Settings.Training.ComputeModelEvidence = computeModelEvidence;
int currentIterationCount = 0;
classifier.IterationChanged += (sender, eventArgs) => { currentIterationCount = eventArgs.CompletedIterationCount; };
// Train classifier
var stopWatch = new Stopwatch();
stopWatch.Start();
classifier.Train(trainingSet);
stopWatch.Stop();
// Produce predictions
var predictions = classifier.PredictDistribution(validationSet).ToList();
var predictedLabels = predictions.Select(
prediction => prediction.Aggregate((aggregate, next) => next.Value > aggregate.Value ? next : aggregate).Key).ToList();
// Iteration count
iterationCounts.Add(currentIterationCount);
// Training time
trainingTime.Add(stopWatch.ElapsedMilliseconds);
// Compute accuracy
accuracy.Add(1 - (evaluator.Evaluate(validationSet, predictedLabels, Metrics.ZeroOneError) / predictions.Count));
// Compute mean negative log probability
negativeLogProbability.Add(evaluator.Evaluate(validationSet, predictions, Metrics.NegativeLogProbability) / predictions.Count);
// Compute M-measure (averaged pairwise AUC)
auc.Add(evaluator.AreaUnderRocCurve(validationSet, predictions));
// Compute log evidence if desired
evidence.Add(computeModelEvidence ? classifier.LogModelEvidence : double.NaN);
// Persist performance metrics
Console.WriteLine(
" Accuracy = {0,5:0.0000} NegLogProb = {1,5:0.0000} AUC = {2,5:0.0000}{3} Iterations = {4} Training time = {5}",
accuracy[fold],
negativeLogProbability[fold],
auc[fold],
computeModelEvidence ? string.Format(" Log evidence = {0,5:0.0000}", evidence[fold]) : string.Empty,
iterationCounts[fold],
BayesPointMachineClassifierModuleUtilities.FormatElapsedTime(trainingTime[fold]));
BayesPointMachineClassifierModuleUtilities.SavePerformanceMetrics(
resultsFile, accuracy, negativeLogProbability, auc, evidence, iterationCounts, trainingTime);
}
return(true);
}
/// <summary>
/// CrossValidate diagnosis
/// </summary>
/// <param name="x"></param>
/// <param name="y"></param>
/// <param name="mapping"></param>
/// <param name="reportFileName"></param>
/// <param name="crossValidationFoldCount"></param>
/// <param name="iterationCount"></param>
/// <param name="computeModelEvidence"></param>
/// <param name="batchCount"></param>
/// <remarks>Adapted from MicrosoftResearch.Infer.Learners</remarks>
public CrossValidateMapped(
Vector[] x,
IList <string> y,
GenericClassifierMapping mapping,
string reportFileName,
int crossValidationFoldCount, //folds
int iterationCount,
bool computeModelEvidence,
int batchCount)
{
Debug.Assert(x != null, "The feature vector must not be null.");
Debug.Assert(y != null, "The targe variable must not be null.");
Debug.Assert(mapping != null, "The classifier map must not be null.");
Debug.Assert(!string.IsNullOrEmpty(reportFileName), "The report file name must not be null/empty.");
Debug.Assert(iterationCount > 0, "The iteration count must be greater than zero.");
Debug.Assert(batchCount > 0, "The batch count must be greater than zero.");
// Shuffle dataset
shuffleVector(x);
// Create evaluator
var evaluatorMapping = mapping.ForEvaluation();
var evaluator = new ClassifierEvaluator <
IList <Vector>, // the type of the instance source,
int, // the type of an instance
IList <string>, // the type of the label source
string>( // the type of a label.
evaluatorMapping);
// Create performance metrics
var accuracy = new List <double>();
var negativeLogProbability = new List <double>();
var auc = new List <double>();
var evidence = new List <double>();
var iterationCounts = new List <double>();
var trainingTime = new List <double>();
// Run cross-validation
int validationSetSize = x.Length / crossValidationFoldCount;
int trainingSetSize = x.Length - validationSetSize;
int validationFoldSetSize = 0;
int trainingFoldSetSize = 0;
Console.WriteLine(
"Running {0}-fold cross-validation", crossValidationFoldCount);
if (validationSetSize == 0 || trainingSetSize == 0)
{
Console.WriteLine("Invalid number of folds");
Console.ReadKey();
System.Environment.Exit(1);
}
for (int fold = 0; fold < crossValidationFoldCount; fold++)
{
// Construct training and validation sets for fold
int validationSetStart = fold * validationSetSize;
int validationSetEnd = (fold + 1 == crossValidationFoldCount)
? x.Length
: (fold + 1) * validationSetSize;
validationFoldSetSize = validationSetEnd - validationSetStart;
trainingFoldSetSize = x.Length - validationFoldSetSize;
Vector[] trainingSet = new Vector[trainingFoldSetSize];
Vector[] validationSet = new Vector[validationFoldSetSize];
IList <string> trainingSetLabels = new List <string>();
IList <string> validationSetLabels = new List <string>();
for (int instance = 0, iv = 0, it = 0; instance < x.Length; instance++)
{
if (validationSetStart <= instance && instance < validationSetEnd)
{
validationSet[iv++] = x[instance];
validationSetLabels.Add(y[instance]);
}
else
{
trainingSet[it++] = x[instance];
trainingSetLabels.Add(y[instance]);
}
}
// Print info
Console.WriteLine(" Fold {0} [validation set instances {1} - {2}]", fold + 1, validationSetStart, validationSetEnd - 1);
// Create classifier
var classifier = BayesPointMachineClassifier.CreateBinaryClassifier(mapping);
classifier.Settings.Training.IterationCount = iterationCount;
classifier.Settings.Training.BatchCount = batchCount;
classifier.Settings.Training.ComputeModelEvidence = computeModelEvidence;
int currentIterationCount = 0;
classifier.IterationChanged += (sender, eventArgs) => { currentIterationCount = eventArgs.CompletedIterationCount; };
// Train classifier
var stopWatch = new Stopwatch();
stopWatch.Start();
classifier.Train(trainingSet, trainingSetLabels);
stopWatch.Stop();
// Produce predictions
IEnumerable <IDictionary <string, double> > predictions =
classifier.PredictDistribution(validationSet);
var predictedLabels = classifier.Predict(validationSet);
// Iteration count
iterationCounts.Add(currentIterationCount);
// Training time
trainingTime.Add(stopWatch.ElapsedMilliseconds);
// Compute accuracy
accuracy.Add(1 - (evaluator.Evaluate(validationSet, validationSetLabels, predictedLabels, Metrics.ZeroOneError) / predictions.Count()));
// Compute mean negative log probability
negativeLogProbability.Add(evaluator.Evaluate(validationSet, validationSetLabels, predictions, Metrics.NegativeLogProbability) / predictions.Count());
// Compute M-measure (averaged pairwise AUC)
auc.Add(evaluator.AreaUnderRocCurve(validationSet, validationSetLabels, predictions));
// Compute log evidence if desired
evidence.Add(computeModelEvidence ? classifier.LogModelEvidence : double.NaN);
// Persist performance metrics
Console.WriteLine(
" Accuracy = {0,5:0.0000} NegLogProb = {1,5:0.0000} AUC = {2,5:0.0000}{3} Iterations = {4} Training time = {5}",
accuracy[fold],
negativeLogProbability[fold],
auc[fold],
computeModelEvidence ? string.Format(" Log evidence = {0,5:0.0000}", evidence[fold]) : string.Empty,
iterationCounts[fold],
FormatElapsedTime(trainingTime[fold]));
SavePerformanceMetrics(
reportFileName, accuracy, negativeLogProbability, auc, evidence, iterationCounts, trainingTime);
}
}