修改随机森林：让特定特征在每次分裂时均被评估

目前主流机器学习库（如scikit-learn、XGBoost等）没有直接提供强制某一特征在每次节点分裂时被纳入评估的参数，但可以通过自定义逻辑实现该需求，以下是两种可行的实现方式：

方式一：手动实现带固定特征的随机森林 #

这种方式直接构建随机森林的训练逻辑，完全控制每棵树的特征选择过程，灵活性更高：

import numpy as np
from sklearn.tree import DecisionTreeClassifier

class CustomRandomForest:
    def __init__(self, n_estimators=100, max_features='sqrt', fixed_feature_idx=15, random_state=None):
        self.n_estimators = n_estimators
        self.max_features = max_features
        self.fixed_feature_idx = fixed_feature_idx
        self.random_state = random_state
        self.trees = []
        self.rng = np.random.RandomState(random_state)
    
    def _get_feature_subset(self, n_features):
        # 解析max_features参数，确定每次选择的特征总数
        if self.max_features == 'sqrt':
            k = int(np.sqrt(n_features))
        elif self.max_features == 'log2':
            k = int(np.log2(n_features))
        elif isinstance(self.max_features, int):
            k = self.max_features
        elif isinstance(self.max_features, float):
            k = int(self.max_features * n_features)
        else:
            k = n_features
        
        # 强制包含指定特征，再从剩余特征中随机补充至k个
        other_features = [i for i in range(n_features) if i != self.fixed_feature_idx]
        if k > 1:
            selected_others = self.rng.choice(other_features, size=k-1, replace=False)
            feature_subset = np.concatenate([[self.fixed_feature_idx], selected_others])
        else:
            feature_subset = np.array([self.fixed_feature_idx])
        return feature_subset
    
    def fit(self, X, y):
        n_samples, n_features = X.shape
        for _ in range(self.n_estimators):
            # 执行bootstrap样本采样
            bootstrap_idx = self.rng.choice(n_samples, size=n_samples, replace=True)
            X_bootstrap = X[bootstrap_idx]
            y_bootstrap = y[bootstrap_idx]
            
            # 获取包含固定特征的子集
            feature_subset = self._get_feature_subset(n_features)
            X_subset = X_bootstrap[:, feature_subset]
            
            # 训练单棵决策树并保存
            tree = DecisionTreeClassifier(random_state=self.rng.randint(0, 10000))
            tree.fit(X_subset, y_bootstrap)
            self.trees.append((tree, feature_subset))
    
    def predict(self, X):
        # 收集所有树的预测结果并投票
        predictions = []
        for tree, feature_subset in self.trees:
            X_subset = X[:, feature_subset]
            predictions.append(tree.predict(X_subset))
        predictions = np.array(predictions)
        return np.apply_along_axis(lambda x: np.bincount(x).argmax(), axis=0, arr=predictions)

使用示例 #

# 假设X是特征矩阵，y是标签
custom_rf = CustomRandomForest(n_estimators=100, fixed_feature_idx=15, random_state=42)
custom_rf.fit(X, y)
predictions = custom_rf.predict(X_test)

方式二：自定义决策树类适配sklearn原生随机森林 #

这种方式通过继承sklearn的DecisionTreeClassifier，重写特征选择逻辑，直接复用sklearn的RandomForestClassifier框架：

from sklearn.tree import DecisionTreeClassifier
from sklearn.utils.validation import check_random_state
from sklearn.ensemble import RandomForestClassifier

class FixedFeatureDecisionTree(DecisionTreeClassifier):
    def __init__(self, fixed_feature_idx=15, **kwargs):
        super().__init__(**kwargs)
        self.fixed_feature_idx = fixed_feature_idx
    
    def _select_features(self, X):
        rng = check_random_state(self.random_state)
        n_features = X.shape[1]
        
        # 解析max_features参数
        if self.max_features == 'sqrt':
            max_features = int(np.sqrt(n_features))
        elif self.max_features == 'log2':
            max_features = int(np.log2(n_features))
        elif isinstance(self.max_features, int):
            max_features = self.max_features
        elif isinstance(self.max_features, float):
            max_features = int(self.max_features * n_features)
        else:
            max_features = n_features
        
        # 构造包含固定特征的候选特征集
        feature_indices = [self.fixed_feature_idx]
        other_indices = [i for i in range(n_features) if i != self.fixed_feature_idx]
        
        if max_features > 1:
            selected_others = rng.choice(other_indices, size=max_features-1, replace=False)
            feature_indices.extend(selected_others)
        
        # 打乱特征顺序，避免固定特征被优先评估
        rng.shuffle(feature_indices)
        return np.array(feature_indices)

使用示例 #

# 用自定义树作为基础估计器构建随机森林
rf = RandomForestClassifier(
    n_estimators=100,
    base_estimator=FixedFeatureDecisionTree(fixed_feature_idx=15),
    random_state=42
)
rf.fit(X, y)
predictions = rf.predict(X_test)

注意事项 #

• 两种方式都保证了指定特征在每棵树的每个节点分裂时都会被纳入候选特征集，但最终是否被选为分裂特征仍由节点分裂的增益计算决定
• 若max_features设置为1，所有节点都会使用指定特征进行分裂，这会导致模型严重过拟合，需谨慎设置参数

内容的提问来源于stack exchange，提问作者Niklas Jacobs