def __init__(self,
n_estimator=10,
learning_rate=0.01,
min_sample=2,
min_gain=0.1,
max_depth=10):
super(GBDTRegressionScratch, self).__init__(n_estimator, learning_rate)
# 回归树损失函数维平方损失
self._loss = SquareLoss()
for _ in range(self._n_estimator):
tree = CARTRegressionScratch(min_sample, min_gain, max_depth)
self._trees.append(tree)
def fit(self, X, y):
"""模型训练"""
# 先对输入标签做one hot编码
y = self._to_one_hot(y)
n_sample, self._n_class = y.shape
# 初始残差为每个类别的平均值
residual_pred = np.full_like(y, np.mean(y, axis=0))
for _ in range(self._n_estimator):
label_trees = []
residual_update = np.zeros_like(residual_pred)
# 每个类别分别学习树
for j in range(self._n_class):
residual_gradient = self._loss.calc_gradient(
y[:, j], residual_pred[:, j])
tree = CARTRegressionScratch(self._min_sample, self._min_gain,
self._max_depth)
# 每棵树以残差为目标进行训练
tree.fit(X, residual_gradient)
label_trees.append(tree)
for i in range(n_sample):
residual_update[i, j] = tree.predict(X[i])
self._trees.append(label_trees)
residual_pred -= self._lr * residual_update