def train(self, num_trees=200, num_features=2, th=0.0):
"""
逐个训练, 采用cart算法
:param num_trees: 树的数量
:param num_features: 每个划分随机抽取的候选特征数量
:param th: 传递给dts.train
:return:
"""
for i in range(num_trees):
# bootstrap有放回抽样
sample_indices = np.random.choice(range(self.m), size=self.m, replace=True)
sampled = [self.rows[i] for i in sample_indices]
# 根据ml_decision_trees.train的设定, 当m=num_features时, 即使sample=True, 也会执行bagging
tree = dts.train_cart(sampled, th=th,
target=self.target, m=num_features, sample=True)
self.trees.append(tree)
print("training complete on {0} trees".format(num_trees))
def train(self, d=1, k=20):
"""
回归问题的提升树算法
:param d: 每棵树的深度
:param k: 树的数量
:return:
"""
def update_y(rows, y_new):
"""
替换数据集rows的y值, 返回一个新的数据集
:param rows: 数据集
:param y_new: 新的y值, list-like
:return:
"""
updated = []
col_id = len(rows[0]) - 1 # 列数
n_rows = len(rows) # 行数
for l in range(n_rows):
r_new = rows[l][:col_id]
r_new.append(y_new[l])
updated.append(r_new)
return updated
for i in range(k):
# 用残差更新训练数据的y值
self.rows_updated = update_y(self.rows_updated, self.resid)
# 训练新的weak learner, 入栈
t = dts.train_cart(self.rows_updated, target="regression", d=d)
self.trees.append(t)
# 更新残差
pred = dts.predict(t, self.rows_updated)
# print(np.array(pred))
# print(self.resid)
self.resid -= np.array(pred)
# 输出当前步骤均方误差
mse = np.dot(self.resid, self.resid) / len(self.resid)
print("training step {0} finished, current mse: {1}".format(str(i), str(mse)))
