如何解决箱线图（boxplot）中unhashable type: 'numpy.ndarray'错误？

错误原因 #

• 传递给sns.boxplot的f1_scores是嵌套数组结构（列表中每个元素是10个交叉验证分数的numpy数组），而seaborn的boxplot需要一维数据对应分组标签，无法直接解析二维嵌套数据。
• 直接将np.arange生成的数组作为参数传递时，seaborn会把它当作分组键，但numpy数组属于不可哈希类型，触发unhashable type错误。
• 你尝试的float(np.arange(.1, .9, .1))完全无效：该np.arange生成的是包含8个元素的数组，无法直接转为单个float，只会引发新的类型错误。

解决方案 #

方法一：使用pandas DataFrame整理数据（推荐） #

seaborn基于DataFrame的接口更直观且不易出错，先将数据整理成长格式DataFrame再传入boxplot：

import pandas as pd
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score

# 假设你的custom_f1函数定义如下（已有则忽略）
def custom_f1(cutoff):
    from sklearn.metrics import make_scorer, f1_score
    def score_func(y_true, y_pred_proba):
        y_pred = (y_pred_proba[:, 1] >= cutoff).astype(int)
        return f1_score(y_true, y_pred)
    return make_scorer(score_func, needs_proba=True)

f1_scores = []
cutoffs = np.arange(.1, .9, .1)

for cutoff in cutoffs:
    clf = RandomForestClassifier(15, random_state=int(cutoff * 10))
    f1_scores.append(cross_val_score(clf, X, y, scoring=custom_f1(cutoff), cv=10))

# 转换为长格式DataFrame
data_rows = []
for cutoff, scores in zip(cutoffs, f1_scores):
    for score in scores:
        data_rows.append({"cutoff": cutoff, "f1_score": score})
df = pd.DataFrame(data_rows)

# 绘制箱线图
sns.boxplot(x="cutoff", y="f1_score", data=df)
sns.despine()
plt.xlabel('cutoff threshold')
plt.ylabel('f1 score')
plt.title('F1 score varies with different cutoff thresholds')
plt.show()

方法二：扁平化数据后直接传递参数 #

若不想使用pandas，可将f1_scores扁平化，同时为每个分数生成对应的cutoff标签：

import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score

# 自定义评分函数（同上）
def custom_f1(cutoff):
    from sklearn.metrics import make_scorer, f1_score
    def score_func(y_true, y_pred_proba):
        y_pred = (y_pred_proba[:, 1] >= cutoff).astype(int)
        return f1_score(y_true, y_pred)
    return make_scorer(score_func, needs_proba=True)

f1_scores = []
cutoffs = np.arange(.1, .9, .1)

for cutoff in cutoffs:
    clf = RandomForestClassifier(15, random_state=int(cutoff * 10))
    f1_scores.append(cross_val_score(clf, X, y, scoring=custom_f1(cutoff), cv=10))

# 扁平化f1分数数组
flat_f1_scores = np.concatenate(f1_scores)
# 生成每个分数对应的cutoff标签（每个cutoff对应10个交叉验证分数）
cutoff_labels = np.repeat(cutoffs, [len(scores) for scores in f1_scores])

# 绘制箱线图
sns.boxplot(x=cutoff_labels, y=flat_f1_scores)
sns.despine()
plt.xlabel('cutoff threshold')
plt.ylabel('f1 score')
plt.title('F1 score varies with different cutoff thresholds')
plt.show()

关键说明 #

核心问题是数据结构不匹配：seaborn的boxplot要求分组变量（x）和数值变量（y）均为一维数组，原代码中的f1_scores是二维嵌套结构，无法被正确解析。不要尝试将多维numpy数组直接转为单个float，这是错误的解决方向。

内容的提问来源于stack exchange，提问作者John