public AdaBoost(string class_variable_value_YES, int T, double threshold = 0)
{
mClassVariableValue_YES = class_variable_value_YES;
mLayer = T;
mWeakClassifiers = new WeakClassifier <T, U> [mLayer];
mAlphaValues = new double[mLayer];
mThreshold = threshold;
}
public override string Predict(T rec)
{
double F_T = 0;
for (int t = 0; t < mLayer; ++t)
{
WeakClassifier <T, U> classifier = mWeakClassifiers[t];
double h = classifier.WeakPredict(rec);
F_T += mAlphaValues[t] * h;
}
return(F_T > 0 ? mClassVariableValue_YES : ClassVariableValue_No);
}
public override void Train(IEnumerable <T> data_store)
{
double h = 0;
double y = 0;
int record_count = 0;
foreach (T rec in data_store)
{
record_count++;
}
double[] weights = new double[record_count];
for (int i = 0; i < record_count; ++i)
{
weights[i] = 1.0 / record_count;
}
int sample_index = 0;
double epsilon = 0;
for (int t = 0; t < mLayer; ++t)
{
// Choose f_t(x)
double min_epsilon = double.MaxValue;
WeakClassifier <T, U> best_classifier = null;
foreach (WeakClassifier <T, U> classifier in mWeakClassifiers)
{
sample_index = 0;
epsilon = 0;
foreach (T rec in data_store)
{
h = classifier.WeakPredict(rec);
y = (mClassVariableValue_YES == rec.Label ? 1 : -1);
epsilon += weights[sample_index] * System.Math.Exp(-h * y);
sample_index++;
}
if (min_epsilon > epsilon)
{
min_epsilon = epsilon;
best_classifier = classifier;
}
}
// Add to ensemble: F_t(x) = F_{t-1}(x) + alpha * h_t(x)
double alpha = 0.5 * System.Math.Log(1 - min_epsilon / min_epsilon);
mAlphaValues[t] = alpha;
mWeakClassifiers[t] = best_classifier;
// Update weights
sample_index = 0;
double weight_sum = 0;
foreach (T rec in data_store)
{
h = best_classifier.WeakPredict(rec);
y = (mClassVariableValue_YES == rec.Label ? 1 : -1);
weights [sample_index] = weights[sample_index] * System.Math.Exp(-h * alpha * y);
weight_sum = weights[sample_index];
sample_index++;
}
for (int i = 0; i < record_count; ++i)
{
weights[sample_index] /= weight_sum;
}
}
}
