当前位置： 首页 > news >正文

企业网站做开放api,中国能源建设集团有限公司官网招聘,网站建设的元素,做网站菠菜什么意思集成学习 不同的算法都可以对解决同一个问题#xff0c;但是可能准确率不同#xff0c;集成学习就是不同算法按照某种组合来解决问题#xff0c;使得准确率提升。 那怎么组合算法呢#xff1f; 自举聚合算法#xff08;bagging#xff09; 顾名思义是 自举聚合 自举…集成学习 不同的算法都可以对解决同一个问题但是可能准确率不同集成学习就是不同算法按照某种组合来解决问题使得准确率提升。 那怎么组合算法呢 自举聚合算法bagging 顾名思义是 自举聚合 自举是指的是自举采样保证随机性允许重复的又放回抽样每次抽与原样本大小相同的样本出来如果进行B次。则有B个数据集然后独立的训练出模型 fx求得平均值 对于低偏差、高方差模型的稳定性有较大提升 随机森林 bagging算法的改进版就是随机森林
from tqdm import tqdm import numpy as np from matplotlib import pyplot as plt from sklearn.datasets import make_classification from sklearn.tree import DecisionTreeClassifier as DTC from sklearn.model_selection import train_test_split# 创建随机数据集 X, y make_classification(n_samples1000, # 数据集大小n_features16, # 特征数即数据维度n_informative5, # 有效特征个数n_redundant2, # 冗余特征个数为有效特征的随机线性组合n_classes2, # 类别数flip_y0.1, # 类别随机的样本个数该值越大分类越困难random_state0 # 随机种子 )print(X.shape) #%% class RandomForest():def init(self, n_trees10, max_featuressqrt):# max_features是DTC的参数表示结点分裂时随机采样的特征个数# sqrt代表取全部特征的平方根None代表取全部特征log2代表取全部特征的对数self.n_trees n_treesself.oob_score 0self.trees [DTC(max_featuresmax_features)for _ in range(n_trees)]# 用X和y训练模型def fit(self, X, y):n_samples, n_features X.shapeself.n_classes np.unique(y).shape[0] # 集成模型的预测累加单个模型预测的分类概率再取较大值作为最终分类ensemble np.zeros((n_samples, self.n_classes))for tree in self.trees:# 自举采样该采样允许重复idx np.random.randint(0, n_samples, n_samples)# 没有被采到的样本unsampled_mask np.bincount(idx, minlengthn_samples) 0unsampled_idx np.arange(n_samples)[unsampled_mask]# 训练当前决策树tree.fit(X[idx], y[idx])# 累加决策树对OOB样本的预测ensemble[unsampled_idx] tree.predict_proba(X[unsampled_idx])# 计算OOB分数由于是分类任务我们用正确率来衡量self.oob_score np.mean(y np.argmax(ensemble, axis1))# 预测类别def predict(self, X):proba self.predict_proba(X)return np.argmax(proba, axis1)def predict_proba(self, X):# 取所有决策树预测概率的平均ensemble np.mean([tree.predict_proba(X)for tree in self.trees], axis0)return ensemble# 计算正确率def score(self, X, y):return np.mean(y self.predict(X)) #%%

算法测试与可视化

num_trees np.arange(1, 101, 5) np.random.seed(0) plt.figure()# bagging算法 oob_score [] train_score [] with tqdm(num_trees) as pbar:for n_tree in pbar:rf RandomForest(n_treesn_tree, max_featuresNone)rf.fit(X, y)train_score.append(rf.score(X, y))oob_score.append(rf.oob_score)pbar.set_postfix({n_tree: n_tree, train_score: train_score[-1], oob_score: oob_score[-1]}) plt.plot(num_trees, train_score, colorblue, labelbagging_train_score) plt.plot(num_trees, oob_score, colorblue, linestyle-., labelbagging_oob_score)# 随机森林算法 oob_score [] train_score [] with tqdm(num_trees) as pbar:for n_tree in pbar:rf RandomForest(n_treesn_tree, max_featuressqrt)rf.fit(X, y)train_score.append(rf.score(X, y))oob_score.append(rf.oob_score)pbar.set_postfix({n_tree: n_tree, train_score: train_score[-1], oob_score: oob_score[-1]}) plt.plot(num_trees, train_score, colorred, linestyle–, labelrandom_forest_train_score) plt.plot(num_trees, oob_score, colorred, linestyle:, labelrandom_forest_oob_score)plt.ylabel(Score) plt.xlabel(Number of trees) plt.legend() plt.show()提升算法 提升算法是另一种集成学习的框架思路是利用当前模型的偏差来调整训练数据的权重
适应提升 from sklearn.ensemble import AdaBoostClassifier

初始化stump

stump DTC(max_depth1, min_samples_leaf1, random_state0)# 弱分类器个数 M np.arange(1, 101, 5) bg_score [] rf_score [] dsc_ada_score [] real_ada_score [] plt.figure()with tqdm(M) as pbar:for m in pbar:# bagging算法bc BaggingClassifier(estimatorstump, n_estimatorsm, random_state0)bc.fit(X_train, y_train)bg_score.append(bc.score(X_test, y_test))# 随机森林算法rfc RandomForestClassifier(n_estimatorsm, max_depth1, min_samples_leaf1, random_state0)rfc.fit(X_train, y_train)rf_score.append(rfc.score(X_test, y_test))# 离散 AdaBoostSAMME是分步加性模型stepwise additive model的缩写dsc_adaboost AdaBoostClassifier(estimatorstump, n_estimatorsm, algorithmSAMME, random_state0)dsc_adaboost.fit(X_train, y_train)dsc_ada_score.append(dsc_adaboost.score(X_test, y_test))# 实 AdaBoostSAMME.R表示弱分类器输出实数real_adaboost AdaBoostClassifier(estimatorstump, n_estimatorsm, algorithmSAMME.R, random_state0)real_adaboost.fit(X_train, y_train)real_ada_score.append(real_adaboost.score(X_test, y_test))# 绘图 plt.plot(M, bg_score, colorblue, labelBagging) plt.plot(M, rf_score, colorred, ls–, labelRandom Forest) plt.plot(M, dsc_ada_score, colorgreen, ls-., labelDiscrete AdaBoost) plt.plot(M, real_ada_score, colorpurple, ls:, labelReal AdaBoost) plt.xlabel(Number of trees) plt.ylabel(Test score) plt.legend() plt.tight_layout() plt.savefig(output_26_1.png) plt.savefig(output_26_1.pdf) plt.show() #%%GBDT算法 GBDT算法中应用广泛的是XGBoost其在损失函数中添加与决策树复杂度相关的正则化约束防止单个弱学习发生过拟合现象。

安装并导入xgboost库

!pip install xgboost import xgboost as xgb from sklearn.datasets import make_friedman1 from sklearn.neighbors import KNeighborsRegressor from sklearn.linear_model import LinearRegression from sklearn.tree import DecisionTreeRegressor from sklearn.ensemble import BaggingRegressor, RandomForestRegressor, \StackingRegressor, AdaBoostRegressor# 生成回归数据集 reg_X, reg_y make_friedman1(n_samples2000, # 样本数目n_features100, # 特征数目noise0.5, # 噪声的标准差random_state0 # 随机种子 )# 划分训练集与测试集 reg_X_train, reg_X_test, reg_y_train, reg_y_test \train_test_split(reg_X, reg_y, test_size0.2, random_state0) #%% def rmse(regressor):# 计算regressor在测试集上的RMSEy_pred regressor.predict(reg_X_test)return np.sqrt(np.mean((y_pred - reg_y_test) ** 2))# XGBoost回归树 xgbr xgb.XGBRegressor(n_estimators100, # 弱分类器数目max_depth1, # 决策树最大深度learning_rate0.5, # 学习率gamma0.0, # 对决策树叶结点数目的惩罚系数当弱分类器为stump时不起作用reg_lambda0.1, # L2正则化系数subsample0.5, # 与随机森林类似表示采样特征的比例objectivereg:squarederror, # MSE损失函数eval_metricrmse, # 用RMSE作为评价指标random_state0 # 随机种子 )xgbr.fit(reg_X_train, reg_y_train) print(fXGBoost{rmse(xgbr):.3f})# KNN回归 knnr KNeighborsRegressor(n_neighbors5).fit(reg_X_train, reg_y_train) print(fKNN{rmse(knnr):.3f})# 线性回归 lnr LinearRegression().fit(reg_X_train, reg_y_train) print(f线性回归{rmse(lnr):.3f})# bagging stump_reg DecisionTreeRegressor(max_depth1, min_samples_leaf1, random_state0) bcr BaggingRegressor(estimatorstump_reg, n_estimators100, random_state0) bcr.fit(reg_X_train, reg_y_train) print(fBagging{rmse(bcr):.3f})# 随机森林 rfr RandomForestRegressor(n_estimators100, max_depth1, max_featuressqrt, random_state0) rfr.fit(reg_X_train, reg_y_train) print(f随机森林{rmse(rfr):.3f})# 堆垛默认元学习器为带L2正则化约束的线性回归 stkr StackingRegressor(estimators[(knn, knnr), (ln, lnr), (rf, rfr) ]) stkr.fit(reg_X_train, reg_y_train) print(fStacking{rmse(stkr):.3f})# 带有输入特征的堆垛 stkr_pt StackingRegressor(estimators[(knn, knnr), (ln, lnr), (rf, rfr) ], passthroughTrue) stkr_pt.fit(reg_X_train, reg_y_train) print(f带输入特征的Stacking{rmse(stkr_pt):.3f})# AdaBoost回归型AdaBoost只有连续型没有离散型 abr AdaBoostRegressor(estimatorstump_reg, n_estimators100, learning_rate1.5, losssquare, random_state0) abr.fit(reg_X_train, reg_y_train)