机器学习算法原理实现——xgboost，核心是加入了正则化和损失函数二阶泰勒展开-526互联

先看总的图：

本质上就是在传统gbdt的决策树基础上加入了正则化防止过拟合，以及为了让损失函数求解更方便，加入了泰勒展开，这样计算损失函数更方便了（除了决策树代码有差别，其他都是gbdt一样，本文仅实现xgboost的决策树）。如下：

再解释各个步骤：

。。。

让gpt来汇总下：

好了，我们直接写下实现代码：

import numpy as np

class XGBoostDecisionTree:
    def __init__(self, max_depth=3):
        self.max_depth = max_depth
        
    def fit(self, X, y):
        # X是特征数据，y的第一列是伪残差，第二列是hessians
        self.tree = self._grow_tree(X, y, depth=0)
        
    def _gain(self, gradients, hessians):
        return np.square(gradients.sum()) / (hessians.sum() + 1e-9)
    
    def _split(self, X, y):
        best_gain = 0
        best_split = None
        best_left_y, best_right_y = None, None
        gradients, hessians = y[:, 0], y[:, 1]
        total_gain = self._gain(gradients, hessians)
        for i in range(X.shape[1]):
            thresholds = np.unique(X[:, i])
            for thresh in thresholds:
                left_mask = X[:, i] < thresh
                right_mask = ~left_mask
                left_gradients = gradients[left_mask]
                right_gradients = gradients[right_mask]
                left_hessians = hessians[left_mask]
                right_hessians = hessians[right_mask]
                left_gain = self._gain(left_gradients, left_hessians)
                right_gain = self._gain(right_gradients, right_hessians)
                gain = left_gain + right_gain - total_gain
                if gain > best_gain:
                    best_gain = gain
                    best_split = (i, thresh)
                    best_left_y = y[left_mask]
                    best_right_y = y[right_mask]
        return best_gain, best_split, best_left_y, best_right_y
    
    def _grow_tree(self, X, y, depth):
        gradients, hessians = y[:, 0], y[:, 1]
        predicted_value = -gradients.sum() / (hessians.sum() + 1e-9)
        node = {"value": predicted_value}
        if depth < self.max_depth:
            gain, split, left_y, right_y = self._split(X, y)
            if gain > 0:  # 只有当增益大于0时我们才真正地进行分裂
                feature_idx, threshold = split
                left_mask = X[:, feature_idx] < threshold
                node["feature_idx"] = feature_idx
                node["threshold"] = threshold
                node["left"] = self._grow_tree(X[left_mask], left_y, depth+1)
                node["right"] = self._grow_tree(X[~left_mask], right_y, depth+1)
        return node
    
    def predict(self, X):
        return np.array([self._predict_single(x, self.tree) for x in X])
    
    def _predict_single(self, x, node):
        if "feature_idx" in node:
            if x[node["feature_idx"]] < node["threshold"]:
                return self._predict_single(x, node["left"])
            else:
                return self._predict_single(x, node["right"])
        return node["value"]

import matplotlib.pyplot as plt

X = np.linspace(0, 10, 100)[:, np.newaxis]
y_true = np.sin(X).ravel() + np.random.normal(0, 0.1, X.shape[0])
base_pred = np.ones_like(y_true) * y_true.mean()
pseudo_residuals = -2 * (y_true - base_pred)
y = np.c_[pseudo_residuals, np.ones_like(y_true)]

model = XGBoostDecisionTree(max_depth=8)
model.fit(X, y)
predictions = model.predict(X) + base_pred  # 加上基学习器的预测

plt.scatter(X, y_true, label='True values', color='b')
plt.plot(X, predictions, label='XGBoost Decision Tree', color='r')
plt.legend()
plt.title('XGBoost Decision Tree Regression')
plt.xlabel('X')
plt.ylabel('y')
plt.show()

效果图：