一.简介
xgboost分类分两种情况，二分类和多分类：
（1） 二分类的思路与logistic回归一样，先对线性函数套一个sigmoid函数，然后再求交叉熵作为损失函数，所以只需要一组回归树并可实现；
（2）而多分类的实现，思路同gbm_classifier一样，即同时训练多组回归树，每一组代表一个class，然后对其进行softmax操作，然后再求交叉熵做为损失函数
下面对多分类的情况再推一次损失函数、一阶导、二阶导：
softmax转换：
[softmax(y^{hat})=softmax([y_1^{hat},y_2^{hat},...,y_n^{hat}])=frac{1}{sum_{i=1}^n e^{y_i^{hat}}}[e^{y_1^{hat}},e^{y_2^{hat}},...,e^{y_n^{hat}}]
]
交叉熵：
[cross\_entropy(y,p)=-sum_{i=1}^n y_ilog p_i
]
将(p_i)替换为(frac{e^{y_i^{hat}}}{sum_{i=1}^n e^{y_i^{hat}}})，得到损失函数如下：
[L(y^{hat},y)=-sum_{i=1}^n y_ilog frac{e^{y_i^{hat}}}{sum_{j=1}^n e^{x_j^{hat}}}\
=-sum_{i=1}^n y_i(y_i^{hat}-logsum_{j=1}^n e^{y_j^{hat}})\
=logsum_{i=1}^n e^{y_i^{hat}}-sum_{i=1}^ny_iy_i^{hat}（由于是onehot展开，所以sum_{i=1}^n y_i=1）
]
所以一阶导：
[frac{partial L(y^{hat},y)}{partial y^{hat}}=softmax([y_1^{hat},y_2^{hat},...,y_n^{hat}])-[y_1,y_2,...,y_n]\
=softmax(y^{hat})-y
]
二阶导：
[frac{partial^2 L(y^{hat},y)}{partial {y^{hat}}^2}=softmax(y^{hat})(1-softmax(y^{hat}))
]
二.代码实现
import os
os.chdir('../')
from ml_models.ensemble import XGBoostBaseTree
from ml_models import utils
import copy
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
"""
xgboost分类树的实现，封装到ml_models.ensemble
"""
class XGBoostClassifier(object):
def __init__(self, base_estimator=None, n_estimators=10, learning_rate=1.0):
"""
:param base_estimator: 基学习器
:param n_estimators: 基学习器迭代数量
:param learning_rate: 学习率，降低后续基学习器的权重，避免过拟合
"""
self.base_estimator = base_estimator
self.n_estimators = n_estimators
self.learning_rate = learning_rate
if self.base_estimator is None:
self.base_estimator = XGBoostBaseTree()
# 同质分类器
if type(base_estimator) != list:
estimator = self.base_estimator
self.base_estimator = [copy.deepcopy(estimator) for _ in range(0, self.n_estimators)]
# 异质分类器
else:
self.n_estimators = len(self.base_estimator)
# 扩展class_num组分类器
self.expand_base_estimators = []
def fit(self, x, y):
# 将y转one-hot编码
class_num = np.amax(y) + 1
y_cate = np.zeros(shape=(len(y), class_num))
y_cate[np.arange(len(y)), y] = 1
# 扩展分类器
self.expand_base_estimators = [copy.deepcopy(self.base_estimator) for _ in range(class_num)]
# 第一个模型假设预测为0
y_pred_score_ = np.zeros(shape=(x.shape[0], class_num))
# 计算一阶、二阶导数
g = utils.softmax(y_pred_score_) - y_cate
h = utils.softmax(y_pred_score_) * (1 - utils.softmax(y_pred_score_))
# 训练后续模型
for index in range(0, self.n_estimators):
y_pred_score = []
for class_index in range(0, class_num):
self.expand_base_estimators[class_index][index].fit(x, g[:, class_index], h[:, class_index])
y_pred_score.append(self.expand_base_estimators[class_index][index].predict(x))
y_pred_score_ += np.c_[y_pred_score].T * self.learning_rate
g = utils.softmax(y_pred_score_) - y_cate
h = utils.softmax(y_pred_score_) * (1 - utils.softmax(y_pred_score_))
def predict_proba(self, x):
# TODO:并行优化
y_pred_score = []
for class_index in range(0, len(self.expand_base_estimators)):
estimator_of_index = self.expand_base_estimators[class_index]
y_pred_score.append(
np.sum(
[estimator_of_index[0].predict(x)] +
[self.learning_rate * estimator_of_index[i].predict(x) for i in
range(1, self.n_estimators - 1)] +
[estimator_of_index[self.n_estimators - 1].predict(x)]
, axis=0)
)
return utils.softmax(np.c_[y_pred_score].T)
def predict(self, x):
return np.argmax(self.predict_proba(x), axis=1)
#造伪数据
from sklearn.datasets import make_classification
data, target = make_classification(n_samples=100, n_features=2, n_classes=2, n_informative=1, n_redundant=0,
n_repeated=0, n_clusters_per_class=1, class_sep=.5,random_state=21)
classifier = XGBoostClassifier()
classifier.fit(data, target)
utils.plot_decision_function(data, target, classifier)
