def updateDistribution(self, tree, dataMatrix, mislabels, distribution, beta): probs = tree.predictProbabilities(dataMatrix.iloc[mislabels[:,0]]) probsIncorrect = np.choose(mislabels[:,1], probs.T) probsCorrect = np.choose(mislabels[:,2], probs.T) power = (0.5 * sum(distribution * (1 + probsCorrect - probsIncorrect))) distribution = distribution * (np.power(beta,power)) distribution = helpers.toProbDistribution(distribution) return distribution
def trainBoosted(self, dataMatrix, classVectors):
mislabels = self.getMislabelSet(np.arange(
dataMatrix.shape[0]), classVectors)
distribution = np.ones(mislabels.shape[0])
distribution = helpers.toProbDistribution(distribution)
numToBag = int(mislabels.shape[0])
for tree in self.trees:
mislabelIndices = helpers.bag(np.arange(mislabels.shape[0]),
numToBag, distribution)
tree.train(dataMatrix.iloc[mislabels[mislabelIndices][:,0]],
classVectors.iloc[mislabels[mislabelIndices][:,0]])
loss = self.calculatePseudoLoss(tree, dataMatrix,
mislabels, distribution)
beta = loss/(1-loss)
self.boostedBeta.append(beta)
distribution = self.updateDistribution(tree, dataMatrix, mislabels,
distribution, beta)
