GBDT之GradientBoostingClassifier源码分析

内容导读

互联网集市收集整理的这篇技术教程文章主要介绍了GBDT之GradientBoostingClassifier源码分析，小编现在分享给大家，供广大互联网技能从业者学习和参考。文章包含3328字，纯文字阅读大概需要5分钟。

内容图文

GradientBoostingClassifier

import pandas as pd
import numpy as np
import math
from sklearn.ensemble import GradientBoostingClassifier
df = pd.DataFrame([[1,-1],[2,-1],[3,-1],[4,1],[5,1],
                   [6,-1],[7,-1],[8,-1],[9,1],[10,1]])
X = df.iloc[:,[0]]
Y = df.iloc[:,-1]
model = GradientBoostingClassifier(n_estimators=20, learning_rate=1.0,
                                 max_depth=1, random_state=0)
model.fit(X, Y)
print(model.predict(X))

模型初始化
第1轮加法模型不在预测为均值
def fit(self, X, y, sample_weight=None):
    # pre-cond: pos, neg are encoded as 1, 0
    if sample_weight is None:
        pos = np.sum(y)
        neg = y.shape[0] - pos
    else:
        pos = np.sum(sample_weight * y)
        neg = np.sum(sample_weight * (1 - y))

    if neg == 0 or pos == 0:
        raise ValueError('y contains non binary labels.')
    self.prior = self.scale * np.log(pos / neg)

计算第1轮加法模型下损失函数的负梯度（残差）

计算负梯度的公式 expit(x) = 1/(1+exp(-x))

def negative_gradient(self, y, pred, **kargs):
    """Compute the residual (= negative gradient). """
    return y - expit(pred.ravel())

调整叶子结点
使用第1轮加法模型的负梯度拟合一棵回归树，但这里要调整叶子结点，每个叶子结点的输出值依赖于选用的损失函数

def update_terminal_regions(self, tree, X, y, residual, y_pred,
                            sample_weight, sample_mask,
                            learning_rate=1.0, k=0):
    """Update the terminal regions (=leaves) of the given tree and
    updates the current predictions of the model. Traverses tree
    and invokes template method `_update_terminal_region`.

    Parameters
    ----------
    tree : tree.Tree
        The tree object.
    X : ndarray, shape=(n, m)
        The data array.
    y : ndarray, shape=(n,)
        The target labels.
    residual : ndarray, shape=(n,)
        The residuals (usually the negative gradient).
    y_pred : ndarray, shape=(n,)
        The predictions.
    sample_weight : ndarray, shape=(n,)
        The weight of each sample.
    sample_mask : ndarray, shape=(n,)
        The sample mask to be used.
    learning_rate : float, default=0.1
        learning rate shrinks the contribution of each tree by
         ``learning_rate``.
    k : int, default 0
        The index of the estimator being updated.

    """
    # compute leaf for each sample in ``X``.
    terminal_regions = tree.apply(X)

    # mask all which are not in sample mask.
    masked_terminal_regions = terminal_regions.copy()
    masked_terminal_regions[~sample_mask] = -1

    # update each leaf (= perform line search)
    for leaf in np.where(tree.children_left == TREE_LEAF)[0]:
        self._update_terminal_region(tree, masked_terminal_regions,
                                     leaf, X, y, residual,
                                     y_pred[:, k], sample_weight)

    # update predictions (both in-bag and out-of-bag)
    y_pred[:, k] += (learning_rate
                     * tree.value[:, 0, 0].take(terminal_regions, axis=0))


numerator = np.sum(sample_weight * residual)
denominator = np.sum(sample_weight * (y - residual) * (1 - y + residual))

可视化

[[1,-1],[2,-1],[3,-1],[4,1],[5,1],[6,-1],[7,-1],[8,-1],[9,1],[10,1]]
round1
第1轮加法模型不是预测为均值了
self.prior = self.scale * np.log(pos / neg)
pos 是1类数量 neg是-1类数量
np.log(4/6)=-0.4054

计算负梯度，程序会自动将-1类变为0
a = np.log(4/6)
print(0-1/(1+np.exp(-a))) -0.4
return y - expit(pred.ravel())

round2
使用第1轮加法模型下的负梯度拟合一棵回归树，这里就需要对叶子结点进行调整了

前8个样本被分在了左叶结点，后2个样本被分在了右叶结点
以左叶结点为例，该结点最终的输出值的计算如下：

分子/分母= -0.625，这里的学习率=1

加入加法模型后
-0.4054-0.625=-1.0304

round3

round4

…
round20

预测

score = self.decision_function(X)  封装
decisions = self.loss_._score_to_decision(score) 输出类别
return self.classes_.take(decisions, axis=0)

内容总结

以上是互联网集市为您收集整理的GBDT之GradientBoostingClassifier源码分析全部内容，希望文章能够帮你解决GBDT之GradientBoostingClassifier源码分析所遇到的程序开发问题。 如果觉得互联网集市技术教程内容还不错，欢迎将互联网集市网站推荐给程序员好友。

内容备注

版权声明：本文内容由互联网用户自发贡献，该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容， 请发送邮件至 gblab@vip.qq.com 举报，一经查实，本站将立刻删除。

内容手机端

---