实战项目：房价预测
用 Scikit-learn 走通完整的机器学习流程，从数据清洗到模型调优。
数据集
使用 Kaggle 的房价预测竞赛数据，80 个特征，预测房屋售价：
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
# 加载数据
train = pd.read_csv('train.csv')
test = pd.read_csv('test.csv')
print(f"训练集: {train.shape}")
print(f"测试集: {test.shape}")
第一步：探索性数据分析（EDA）
目标变量分布
plt.figure(figsize=(10, 4))
plt.subplot(1, 2, 1)
sns.histplot(train['SalePrice'], bins=50)
plt.title('SalePrice Distribution')
# SalePrice 通常右偏，取对数使其更接近正态
plt.subplot(1, 2, 2)
sns.histplot(np.log1p(train['SalePrice']), bins=50)
plt.title('log(SalePrice) Distribution')
缺失值分析
# 查看缺失比例
missing = train.isnull().sum().sort_values(ascending=False)
missing_pct = missing[missing > 0] / len(train) * 100
print(missing_pct.head(20))
# 规则：
# - 缺失 > 50% → 删除该特征
# - 缺失 < 5% → 填充（数值用中位数，类别用众数）
# - 缺失在 5-50% → 创建"是否缺失"的指示特征
相关性分析
# 目标与数值特征的相关性
numeric_cols = train.select_dtypes(include=[np.number]).columns
corr = train[numeric_cols].corr()['SalePrice'].sort_values(ascending=False)
print(corr.head(15))
print(corr.tail(5))
第二步：数据清洗
def clean_data(df, is_train=True):
data = df.copy()
# ---- 处理缺失值 ----
# 数值特征：填中位数
num_cols = data.select_dtypes(include=[np.number]).columns
for col in num_cols:
if data[col].isnull().sum() > 0:
data[col] = data[col].fillna(data[col].median())
# 类别特征：填众数
cat_cols = data.select_dtypes(include=['object']).columns
for col in cat_cols:
if data[col].isnull().sum() > 0:
data[col] = data[col].fillna(data[col].mode()[0])
# ---- 目标变量取对数 ----
if is_train:
data['SalePrice'] = np.log1p(data['SalePrice'])
return data
train_clean = clean_data(train, is_train=True)
test_clean = clean_data(test, is_train=False)
第三步：特征工程
# 合并训练和测试集做统一编码
all_data = pd.concat([train_clean, test_clean], sort=False)
# 类别特征做 One-Hot / Label 编码
# 高基数特征（如 Neighborhood）用 Target Encoding
from sklearn.preprocessing import LabelEncoder
cat_cols = all_data.select_dtypes(include=['object']).columns
for col in cat_cols:
le = LabelEncoder()
all_data[col] = le.fit_transform(all_data[col].astype(str))
# ---- 构造新特征 ----
# 总面积
all_data['TotalSF'] = (
all_data['TotalBsmtSF'] +
all_data['1stFlrSF'] +
all_data['2ndFlrSF']
)
# 房屋年龄（售出年份 - 建造年份）
all_data['HouseAge'] = all_data['YrSold'] - all_data['YearBuilt']
# 是否有地下室
all_data['HasBsmt'] = (all_data['TotalBsmtSF'] > 0).astype(int)
# ---- 分离回训练/测试 ----
feat_cols = [c for c in all_data.columns
if c not in ['Id', 'SalePrice']]
X = all_data[:len(train_clean)][feat_cols]
y = train_clean['SalePrice']
X_test = all_data[len(train_clean):][feat_cols]
print(f"特征维度: {X.shape}")
第四步：建模
基线模型
from sklearn.linear_model import LinearRegression, Ridge
from sklearn.model_selection import cross_val_score
def rmse_cv(model, X, y, n_folds=5):
"""交叉验证的 RMSE"""
scores = np.sqrt(-cross_val_score(
model, X, y,
scoring='neg_mean_squared_error',
cv=n_folds, n_jobs=-1,
))
return scores.mean(), scores.std()
# 基线：线性回归
lr = LinearRegression()
mean, std = rmse_cv(lr, X, y)
print(f"LR RMSE: {mean:.5f} (±{std:.5f})")
# Ridge 回归
ridge = Ridge(alpha=10)
mean, std = rmse_cv(ridge, X, y)
print(f"Ridge RMSE: {mean:.5f} (±{std:.5f})")
集成模型
from sklearn.ensemble import RandomForestRegressor, GradientBoostingRegressor
# 随机森林
rf = RandomForestRegressor(
n_estimators=300, max_depth=15,
min_samples_split=5, n_jobs=-1, random_state=42,
)
mean, std = rmse_cv(rf, X, y)
print(f"RF RMSE: {mean:.5f} (±{std:.5f})")
# Gradient Boosting
gbr = GradientBoostingRegressor(
n_estimators=500, learning_rate=0.05,
max_depth=4, min_samples_split=5, random_state=42,
)
mean, std = rmse_cv(gbr, X, y)
print(f"GBR RMSE: {mean:.5f} (±{std:.5f})")
第五步：参数调优
from sklearn.model_selection import RandomizedSearchCV
from scipy.stats import randint
param_dist = {
'n_estimators': randint(200, 600),
'max_depth': randint(3, 20),
'min_samples_split': randint(2, 20),
'min_samples_leaf': randint(1, 10),
}
rf = RandomForestRegressor(n_jobs=-1, random_state=42)
search = RandomizedSearchCV(
rf, param_dist, n_iter=30,
cv=5, scoring='neg_mean_squared_error',
n_jobs=-1, random_state=42,
)
search.fit(X, y)
print(f"最佳参数: {search.best_params_}")
print(f"最佳 RMSE: {np.sqrt(-search.best_score_):.5f}")
best_model = search.best_estimator_
第六步：特征重要性分析
importances = best_model.feature_importances_
indices = np.argsort(importances)[::-1]
# 前 15 个最重要的特征
for i in range(15):
print(f"{i+1}. {feat_cols[indices[i]]}: {importances[indices[i]]:.4f}")
# 可视化
plt.figure(figsize=(10, 8))
plt.barh(range(15), importances[indices[:15]][::-1])
plt.yticks(range(15), [feat_cols[i] for i in indices[:15]][::-1])
plt.xlabel('Feature Importance')
第七步：模型融合
将多个模型的预测做加权平均，通常能获得最后一点提升：
from sklearn.ensemble import VotingRegressor
ensemble = VotingRegressor([
('ridge', Ridge(alpha=10)),
('rf', RandomForestRegressor(n_estimators=300, max_depth=12)),
('gbr', GradientBoostingRegressor(n_estimators=500, learning_rate=0.05)),
])
mean, std = rmse_cv(ensemble, X, y)
print(f"Ensemble RMSE: {mean:.5f} (±{std:.5f})")
总结
一个完整的 ML 项目流程：
EDA → 数据清洗 → 特征工程 → 基线模型 → 集成模型 → 调优 → 模型融合
几点经验：
- 不要跳过 EDA — 理解数据比调参更重要
- 先搭基线（线性回归/逻辑回归），确保 pipeline 跑通再上复杂模型
- 特征工程决定上限 — 算法的提升空间往往不如好的特征
- 交叉验证比单次划分可靠 — 避免意外过拟合
- 模型融合是锦上添花 — 单模型调好之后再考虑