Python实现ID3算法计算信息增益返回Series而非数值问题

返回pandas.Series而非单个数值的核心原因是维度操作不匹配，常见触发场景有两种：

• 计算条件熵时未指定仅对标签列做聚合统计，pandas自动对所有列执行分组计数，返回携带所有列名的序列
• 调用增益计算函数时传入了整个DataFrame而非单个特征列，或是用apply全局遍历后未做单列索引，直接输出了全列的计算结果

核心修正逻辑 #

1. 条件熵计算时，分组后仅针对标签列做统计，避免返回多列结果
2. 明确传入单个特征列计算增益，如需遍历全列再额外做循环处理
3. 所有聚合操作后优先提取数值，避免携带pandas的索引结构

可运行修正代码 #

import pandas as pd
import numpy as np

# 计算数据集经验熵
def calc_root_entropy(df, label_col='label'):
    label_cnt = df[label_col].value_counts().values
    total = len(df)
    ent = 0.0
    for cnt in label_cnt:
        prob = cnt / total
        ent -= prob * np.log2(prob)
    return ent

# 计算单个特征的条件熵
def calc_condition_entropy(df, feature_col, label_col='label'):
    total = len(df)
    # 仅对标签列做分组统计，避免多列返回
    feature_group = df.groupby(feature_col)[label_col].count()
    cond_ent = 0.0
    for feat_val, cnt in feature_group.items():
        prob = cnt / total
        # 计算当前特征取值分组内的标签熵
        sub_df = df[df[feature_col] == feat_val]
        sub_label_cnt = sub_df[label_col].value_counts().values
        sub_ent = 0.0
        for sub_cnt in sub_label_cnt:
            sub_prob = sub_cnt / cnt
            sub_ent -= sub_prob * np.log2(sub_prob)
        cond_ent += prob * sub_ent
    return cond_ent

# 计算单个特征的信息增益，直接返回float标量
def calc_single_feature_gain(df, feature_col, label_col='label'):
    root_ent = calc_root_entropy(df, label_col)
    return root_ent - calc_condition_entropy(df, feature_col, label_col)

调用示例 #

# 计算单个特征的增益，返回单个数值
age_gain = calc_single_feature_gain(train_df, 'age')
print(age_gain) # 输出为单个float值

# 遍历所有特征计算增益，按需存储
feature_gain = {}
for col in train_df.columns:
    if col == 'label':
        continue
    feature_gain[col] = calc_single_feature_gain(train_df, col)
print(feature_gain)

如果之前是用apply批量计算得到的Series，直接按列名索引即可拿到单个值：

all_gain = train_df.drop(columns='label').apply(lambda x: calc_single_feature_gain(train_df, x.name))
# 取单个特征的增益
single_gain = all_gain['age']

内容的提问来源于stack exchange，提问作者Evan Gertis