Sklearn中DecisionTreeClassifier的GridSearchCV在Linux/macOS偶现UndefinedMetricWarning

将决策树分类器的网格搜索工作流从Windows迁移至Linux集群后，频繁触发UndefinedMetricWarning警告，但Windows环境下运行无异常。相关信息如下：

运行代码 #

import pandas as pd
from sklearn.tree import DecisionTreeClassifier
from sklearn.metrics import make_scorer,accuracy_score,recall_score,f1_score, precision_score, roc_auc_score
import numpy as np
from sklearn.model_selection import GridSearchCV
data = pd.read_csv("Dtree_test_data.csv")
d_tree = DecisionTreeClassifier()
scoring = {'accuracy' : make_scorer(accuracy_score), 
       'precision' : make_scorer(precision_score, average = 'weighted'),
       'recall' : make_scorer(recall_score, average = 'weighted'), 
       'f1_score' : make_scorer(f1_score, average = 'weighted')}
features = data[["H-Bonds","Mass"]]
labels = data["Class"]
parameters = {'max_depth':np.arange(1,3,1),
              'min_samples_leaf': np.arange(0.01,0.05,0.01),
              'min_impurity_decrease':np.arange(0.01,0.04,0.01),
              'criterion':['gini','entropy']}

dtree = DecisionTreeClassifier()
clf = GridSearchCV(dtree,parameters,verbose=1,n_jobs=12,scoring=scoring,refit=False,return_train_score=True,cv=2)
clf.fit(features,labels)

样本数据 #

H-Bonds  Mass  Class
0        2   123      0
1        3    45      1
2        1   153      2
3        4    90      0
4        6   300      1
5        1    40      2
6        2   200      0
7        3   245      1
8        4    87      2
9        1   126      1

警告信息 #

Fitting 2 folds for each of 48 candidates, totalling 96 fits
/../miniconda3/envs/ml-debug-current/lib/python3.10/site-packages/sklearn/metrics/_classification.py:1334: UndefinedMetricWarning: Precision is ill-defined and being set to 0.0 in labels with no predicted samples. Use zero_division parameter to control this behavior.
_warn_prf(average, modifier, msg_start, len(result))

1. 交叉验证拆分的随机性差异：Windows和Linux环境未固定随机种子，导致GridSearchCV的2折交叉验证拆分结果不一致。你的样本量仅10条，部分拆分后的测试集中，模型可能无法预测到某类标签，触发精度计算的警告；而Windows环境刚好拆分出的测试集都能被模型覆盖所有类别，因此无警告。
2. 加权精度的计算特性：你使用了average='weighted'的精度评分器，当某类标签无预测样本时，该类精度会被强制设为0.0，Scikit-learn会触发此警告。
3. Scikit-learn版本差异：两个环境的Scikit-learn版本可能不同，新版本对未覆盖类别的精度计算警告逻辑更严格，导致Linux环境出现警告而Windows没有。

• 固定全局随机种子：确保跨环境的交叉验证拆分一致，消除随机性影响：

  import random
  random.seed(42)
  np.random.seed(42)
  

• 设置zero_division参数：在精度和F1的评分器中指定无预测样本时的处理方式，直接消除警告：

  'precision' : make_scorer(precision_score, average='weighted', zero_division=1),
  'f1_score' : make_scorer(f1_score, average='weighted', zero_division=1)
  

• 改用分层交叉验证：针对小样本的多分类任务，使用StratifiedKFold确保拆分后每个折的类别分布与原数据一致，减少极端拆分情况：

  from sklearn.model_selection import StratifiedKFold
  clf = GridSearchCV(dtree, parameters, verbose=1, n_jobs=12, scoring=scoring, refit=False, return_train_score=True, cv=StratifiedKFold(n_splits=2, shuffle=True, random_state=42))
  

内容的提问来源于stack exchange，提问作者doom4