随机森林回归实战：解决NaN指标问题与模型调优指南

1次阅读

共计 2897 个字符，预计需要花费 8 分钟才能阅读完成。

当随机森林回归模型输出 MSE/R²等指标为 NaN 时，通常说明模型未能从数据中学到有效规律。以下是五大常见原因及诊断代码示例：

import numpy as np
import pandas as pd
from sklearn.ensemble import RandomForestRegressor

# 模拟问题数据
data = pd.DataFrame({'feature1': [1, 2, 3, np.nan, 5],
    'feature2': [0, 0, 0, 0, 0],  # 零方差特征
    'target': [1.1, 2.4, 3.0, 4.2, 5.7]
})

# 诊断函数
def check_nan_issues(df: pd.DataFrame) -> dict:
    issues = {'missing_values': df.isnull().sum().to_dict(),
        'zero_variance': df.std()[df.std() == 0].index.tolist(),
        'dtypes': df.dtypes.to_dict()}
    return issues

print(check_nan_issues(data))

关键诊断点：

缺失值（Missing Values）：未处理的 NaN 会破坏决策树分裂
零方差特征（Zero-Variance Features）：如 feature2 所有值相同
数据类型不匹配：分类变量未正确编码
样本量过少：特征数 >> 样本数时容易过拟合
目标变量异常：存在无限大值或全部相同值

from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import OneHotEncoder, StandardScaler

# 定义预处理步骤
numeric_features = ['feature1']
numeric_transformer = Pipeline(steps=[('imputer', SimpleImputer(strategy='median')),
    ('scaler', StandardScaler())
])

categorical_features = ['category_col']
categorical_transformer = Pipeline(steps=[('imputer', SimpleImputer(strategy='constant', fill_value='missing')),
    ('onehot', OneHotEncoder(handle_unknown='ignore'))
])

preprocessor = ColumnTransformer(
    transformers=[('num', numeric_transformer, numeric_features),
        ('cat', categorical_transformer, categorical_features)
    ])

from sklearn.feature_selection import VarianceThreshold

selector = VarianceThreshold(threshold=0.01)  # 移除方差 <0.01 的特征
X_processed = selector.fit_transform(X)

from sklearn.model_selection import GridSearchCV

param_grid = {'n_estimators': [50, 100, 200],
    'max_depth': [None, 10, 20],
    'min_samples_split': [2, 5],
    'min_samples_leaf': [1, 2]
}

model = Pipeline(steps=[('preprocessor', preprocessor),
    ('regressor', RandomForestRegressor())
])

grid_search = GridSearchCV(
    model, param_grid, cv=5,
    scoring='neg_mean_squared_error',
    n_jobs=-1, verbose=2
)
grid_search.fit(X_train, y_train)

关键参数说明：

n_estimators：树的数量，更多树 = 更稳定但计算成本更高
max_depth：控制过拟合，None 表示不限制
min_samples_split：节点分裂最小样本数

解决方案：

# 调整类别权重
class_weight = compute_sample_weight('balanced', y)
model.fit(X, y, sample_weight=class_weight)

解决方法：

设置 n_jobs=1 减少并行度
使用 max_samples 参数限制每棵树使用的样本比例

验证方法：

import matplotlib.pyplot as plt

feat_importances = pd.Series(model.named_steps['regressor'].feature_importances_,
    index=get_feature_names(preprocessor)
)
feat_importances.nlargest(10).plot(kind='barh')
plt.title('Top 10 Feature Importance')
plt.show()

自定义评估函数避免 NaN：

from sklearn.metrics import make_scorer

def safe_r2_score(y_true, y_pred):
    if len(np.unique(y_true)) == 1:
        return 0.0  # 全部预测为均值时返回 0
    return r2_score(y_true, y_pred)

custom_scorer = make_scorer(safe_r2_score)