def __init__(self,
n_estimator=10,
learning_rate=0.01,
min_sample=2,
min_gain=0.1,
max_depth=10):
super(GBDTClassificationScratch,
self).__init__(n_estimator, learning_rate)
"""
min_sample:
当数据集样本数少于min_sample时不再划分
min_gain:
如果划分后收益不能超过该值则不进行划分
对分类树来说基尼指数需要有足够的下降
对回归树来说平方误差要有足够的下降
max_depth:
树的最大高度
"""
self._min_sample = min_sample
self._min_gain = min_gain
self._max_depth = max_depth
# 分类树损失函数维交叉熵损失
self._loss = CrossEntropyLoss()
class GBDTClassificationScratch(GBDTScratch):
def __init__(self,
n_estimator=10,
learning_rate=0.01,
min_sample=2,
min_gain=0.1,
max_depth=10):
super(GBDTClassificationScratch,
self).__init__(n_estimator, learning_rate)
"""
min_sample:
当数据集样本数少于min_sample时不再划分
min_gain:
如果划分后收益不能超过该值则不进行划分
对分类树来说基尼指数需要有足够的下降
对回归树来说平方误差要有足够的下降
max_depth:
树的最大高度
"""
self._min_sample = min_sample
self._min_gain = min_gain
self._max_depth = max_depth
# 分类树损失函数维交叉熵损失
self._loss = CrossEntropyLoss()
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
def predict(self, x):
"""给定输入样本，预测输出"""
y_pred = np.zeros(self._n_class)
for label_trees in self._trees:
for i in range(len(label_trees)):
residual_update = label_trees[i].predict(x)
y_pred[i] -= self._lr * residual_update
# 返回概率值最大的类别，省略了指数计算
return np.argmax(y_pred)
def _to_one_hot(self, y):
"""将离散标签进行one hot编码"""
n_col = np.amax(y) + 1
one_hot = np.zeros((y.shape[0], n_col))
# 将类别所在列置为1
one_hot[np.arange(y.shape[0]), y] = 1
return one_hot