如何基于HistGradientBoostingClassifier绘制决策树？

我有一个已拟合好的HistGradientBoostingClassifier模型（命名为RF_90），想绘制其中一棵或多棵决策树，但没找到原生实现的函数。我能访问TreePredictor对象及其节点，但sklearn.tree.plot_tree只支持DecisionTree类型的对象。我试了下面的代码：

from sklearn.tree import plot_tree

plot_tree(RF_90._predictors[0][0])

结果触发了错误：

InvalidParameterError: The 'decision_tree' parameter of plot_tree must
be an instance of 'sklearn.tree._classes.DecisionTreeClassifier' or an
instance of 'sklearn.tree._classes.DecisionTreeRegressor'. Got
<sklearn.ensemble._hist_gradient_boosting.predictor.TreePredictor
object at 0x7f676ebf0310> instead.

TreePredictor是HistGradientBoosting系列模型的内部预测器，并不兼容plot_tree的输入要求，你可以通过以下两种方式实现树的可视化：

方法1：手动解析TreePredictor节点绘制 #

TreePredictor的nodes_属性存储了树的核心节点数据（特征索引、阈值、子节点索引、样本数等），可以遍历这些数据，用matplotlib手动绘制树结构：

import matplotlib.pyplot as plt
from matplotlib.pylab import rcParams

rcParams['figure.figsize'] = 15, 20

def plot_tree_predictor(tree_predictor, feature_names=None, ax=None):
    if ax is None:
        ax = plt.gca()
    
    nodes = tree_predictor.nodes_
    def draw_node(node_idx, x, y, width):
        node = nodes[node_idx]
        # 绘制叶子节点
        if node.is_leaf:
            ax.text(x, y, f"叶子节点\n输出值: {node.value[0]:.2f}", 
                    ha='center', va='center', bbox=dict(boxstyle='round', fc='white'))
            return
        # 绘制分裂节点
        feature = feature_names[node.feature_idx] if feature_names else f"特征 {node.feature_idx}"
        ax.text(x, y, f"{feature} ≤ {node.threshold:.2f}\n样本数: {node.n_samples}",
                ha='center', va='center', bbox=dict(boxstyle='round', fc='lightblue'))
        # 绘制子节点连线
        left_x = x - width/2
        right_x = x + width/2
        next_y = y - 1
        ax.plot([x, left_x], [y-0.2, next_y+0.2], 'k-')
        ax.plot([x, right_x], [y-0.2, next_y+0.2], 'k-')
        # 递归绘制子节点
        draw_node(node.left_child_idx, left_x, next_y, width/2)
        draw_node(node.right_child_idx, right_x, next_y, width/2)
    
    draw_node(0, 0.5, 1, 0.5)
    ax.axis('off')
    plt.show()

# 使用示例，替换your_feature_names为你的特征列表
plot_tree_predictor(RF_90._predictors[0][0], feature_names=your_feature_names)

方法2：转换为兼容的DecisionTree对象 #

手动构建DecisionTreeClassifier对象，将TreePredictor的节点数据映射到该对象的内部结构，即可直接用plot_tree绘图：

from sklearn.tree import DecisionTreeClassifier, plot_tree

def tree_predictor_to_decision_tree(tree_predictor, feature_names=None):
    dt = DecisionTreeClassifier()
    dt.n_features_in_ = tree_predictor.n_features_
    dt.feature_names_in_ = feature_names if feature_names else [f"特征{i}" for i in range(dt.n_features_in_)]
    
    # 模拟DecisionTree的tree_属性结构
    dt.tree_ = type('MockTree', (), {})()
    nodes = tree_predictor.nodes_
    dt.tree_.node_count = len(nodes)
    dt.tree_.children_left = [node.left_child_idx if not node.is_leaf else -1 for node in nodes]
    dt.tree_.children_right = [node.right_child_idx if not node.is_leaf else -1 for node in nodes]
    dt.tree_.feature = [node.feature_idx if not node.is_leaf else -2 for node in nodes]
    dt.tree_.threshold = [node.threshold if not node.is_leaf else -2 for node in nodes]
    dt.tree_.value = [[node.value] for node in nodes]
    dt.tree_.n_node_samples = [node.n_samples for node in nodes]
    dt.tree_.max_depth = tree_predictor.max_depth_
    
    return dt

# 转换并绘图
dt = tree_predictor_to_decision_tree(RF_90._predictors[0][0], feature_names=your_feature_names)
plot_tree(dt, feature_names=your_feature_names, filled=True)
plt.show()

注意：这种转换依赖sklearn内部属性结构，不同版本可能存在兼容性问题，使用前建议测试。

内容的提问来源于stack exchange，提问作者Alejandro Gonzalez