private TrainingExample OverSampleGeneratively(TrainingSet trainingSet)
{
var featureVector = trainingSet.MinorityFeatureFreq.Select(freq => GenerateFeature(freq, trainingSet.Total)).ToArray();
return new TrainingExample
{
Features = new FeatureVector(featureVector),
ExpectedResult = 1
};
}
public TrainingExample[] OverSample(TrainingExample[] trainingExamples, SamplingParams samplingParams)
{
if (!samplingParams.NeedSampling)
return trainingExamples;
var trainingSet = new TrainingSet(trainingExamples);
var delta = Math.Min(trainingSet.Majority.Length - trainingSet.Minority.Length, trainingSet.Minority.Length * samplingParams.MinorityClassMaxOversampling);
var additionalSamples = Enumerable.Range(0, delta).Select(x => OverSample(trainingSet, samplingParams)).ToArray();
return trainingExamples.Concat(additionalSamples).RandomShuffle().ToArray();
}
private TrainingExample OverSample(TrainingSet trainingSet, SamplingParams samplingParams)
{
switch (samplingParams.Strategy)
{
case SamplingStrategy.Random:
return OverSampleRandomly(trainingSet);
case SamplingStrategy.Generative:
return OverSampleGeneratively(trainingSet);
case SamplingStrategy.Perturbed:
return OverSamplePerturbed(trainingSet);
default:
throw new InvalidOperationException(string.Format("No implementation for strategy {0}", samplingParams.Strategy));
}
}
private TrainingExample OverSamplePerturbed(TrainingSet trainingSet)
{
int total = trainingSet.Total;
var example = trainingSet.Minority[randomProvider.GetInt(0, trainingSet.Minority.Length)].Features.Values;
var featureVector = trainingSet.MinorityFeatureFreq.Select((freq, i) =>
{
var rand = randomProvider.GetInt(0, 100);
if (rand < 70)
return example[i];
if (rand < 80)
return GenerateFeature(freq, total);
return (example[i] + GenerateFeature(freq, total)) / 2;
}).ToArray();
return new TrainingExample
{
Features = new FeatureVector(featureVector),
ExpectedResult = 1
};
}
private TrainingExample OverSampleRandomly(TrainingSet trainingSet)
{
var index = randomProvider.GetInt(0, trainingSet.Minority.Length);
return trainingSet.Minority[index];
}