private void RemoveFeaturesWithConstantValue(AppIdentAcordSource appIdentAcordSource)
{
var matrix = appIdentAcordSource.Samples2D;
var sampleCount = matrix.GetLength(0);
var featureCount = matrix.GetLength(1);
var removedFeatures = 0;
for (var featureIndex = 0; featureIndex < featureCount; featureIndex++)
{
var isConstant = true;
var firstValue = matrix[0, featureIndex];
for (var sampleIndex = 1; sampleIndex < sampleCount; sampleIndex++)
{
if (!(Math.Abs(firstValue - matrix[sampleIndex, featureIndex]) > 0))
{
continue;
}
isConstant = false;
break;
}
if (isConstant)
{
var feature = this.FeatureSelector.SelectedFeatures[featureIndex - removedFeatures];
removedFeatures++;
appIdentAcordSource.FeatureSelector.RemoveFeature(feature);
}
}
}
public RandomForest GetBestRandomForestsWithGridSearch(AppIdentAcordSource appIdentAcordSource, out GridSearchParameterCollection bestParameters, out double minError)
{
// grid search ranges (parameter values)
GridSearchRange[] parameterRanges =
{
new GridSearchRange("trees", new double[]
{
//1,
//3,
//5,
//8,
11,
13,
17,
19,
37
}),
new GridSearchRange("sampleRatio", new[]
{
// 0.7,
0.8,
// 0.9
}),
new GridSearchRange("join", new double[]
{
//25,
//50,
//100,
150,
200,
250,
300
})
};
var samples = appIdentAcordSource.Samples;
var labels = appIdentAcordSource.LabelsAsIntegers;
var decisionVariables = appIdentAcordSource.DecisionVariables;
// instantiate grid search algorithm for a CLF model
var gridSearch = new GridSearch <RandomForest>(parameterRanges)
{
Fitting = delegate(GridSearchParameterCollection parameters, out double error)
{
Console.WriteLine($"{DateTime.Now} RandomForest grid search.");
// Use the parameters to build the model
// Create a new learning algorithm
var rfcModel = CreateRandomForestModel(decisionVariables, parameters, samples, labels);
// Measure the model performance to return as an out parameter
error = new ZeroOneLoss(labels).Loss(rfcModel.Decide(samples));
// Return the current model
return(rfcModel);
}
//,ParallelOptions = new ParallelOptions() { MaxDegreeOfParallelism = 1 }
};
// Compute the grid search to find the best RandomForest model
return(gridSearch.Compute(out bestParameters, out minError));
}
public IEnumerable <double[, ]> ProcessFeatureSelection(AppIdentAcordSource appIdentAcordSource, double trashold)
{
this.FeatureSelector = appIdentAcordSource.FeatureSelector;
this.RemoveFeaturesWithConstantValue(appIdentAcordSource);
var iterationResults = new List <double[, ]>();
Type mostCorrelatedFeature = null;
double correlation = 1;
while (correlation > trashold)
{
if (mostCorrelatedFeature != null)
{
appIdentAcordSource.FeatureSelector.RemoveFeature(mostCorrelatedFeature);
}
var correlationMatrix = this.GetCorrelationMatrix(appIdentAcordSource);
iterationResults.Add(correlationMatrix);
mostCorrelatedFeature = this.GetMostCorellatedFeature(correlationMatrix, out correlation);
}
this.FeatureSelector = null;
return(iterationResults);
}
private double[,] GetCorrelationMatrix(AppIdentAcordSource appIdentAcordSource) => appIdentAcordSource.Samples2D.Correlation();
public CrossValidationResult <RandomForest> GetCrossValidationResultsOfRandomForestModel(AppIdentAcordSource appIdentAcordSource, GridSearchParameterCollection bestParameters, int folds = 10)
{
var samples = appIdentAcordSource.Samples;
var labels = appIdentAcordSource.LabelsAsIntegers;
var decisionVariables = appIdentAcordSource.DecisionVariables;
// Create a new Cross-validation algorithm passing the data set size and the number of folds
var crossvalidation = new CrossValidation <RandomForest>(samples.Length, folds)
{
Fitting = delegate(int k, int[] indicesTrain, int[] indicesValidation)
{
// The fitting function is passing the indices of the original set which
// should be considered training data and the indices of the original set
// which should be considered validation data.
Console.WriteLine($"{DateTime.Now} RandomForest cross validation.");
// Lets now grab the training data:
var trainingInputs = samples.Get(indicesTrain);
var trainingOutputs = labels.Get(indicesTrain);
// And now the validation data:
var validationInputs = samples.Get(indicesValidation);
var validationOutputs = labels.Get(indicesValidation);
// create random forest model with the best parameters from grid search results
var rfcModel = CreateRandomForestModel(decisionVariables, bestParameters, trainingInputs, trainingOutputs);
// compute the training error rate with ZeroOneLoss function
var trainingError = new ZeroOneLoss(trainingOutputs).Loss(rfcModel.Decide(trainingInputs));
// Now we can compute the validation error on the validation data:
var validationError = new ZeroOneLoss(validationOutputs).Loss(rfcModel.Decide(validationInputs));
// Return a new information structure containing the model and the errors achieved.
var tag = new ValidationDataSource(validationInputs, validationOutputs);
return(new CrossValidationValues <RandomForest>(rfcModel, trainingError, validationError)
{
Tag = tag
});
}
};
// Compute the cross-validation
return(crossvalidation.Compute());
}
