我想知道您是否可以帮助我解决我在运行网格搜索中收到的错误。我认为这可能是由于对网格搜索的实际工作方式的误解。
我现在正在运行一个应用程序，我需要网格搜索来评估使用不同评分函数的最佳参数。我使用RandomForestClassifier将一个大型X数据集拟合到一个特征向量Y上，Y是一个0和1的列表。(完全二元)。我的评分函数(MCC)要求预测输入和实际输入完全是二进制的。然而，由于某种原因，我一直得到ValueError: multiclass是不支持的。
我的理解是网格搜索，对数据集进行交叉验证，提出基于交叉验证的预测输入，然后将表征向量和预测插入函数中。由于我的表征向量完全是二进制的，所以我的预测向量也应该是二进制的，这样在评估分数时就不会出现问题。当我运行带有单个定义参数的随机森林(不使用网格搜索)时，将预测数据和特征向量插入MCC评分函数中运行得非常好。所以我有点迷失在如何运行网格搜索会导致任何错误。
数据快照:
print len(X)
print X[0]
print len(Y)
print Y[2990:3000]
17463699
[38.110903683955435, 38.110903683955435, 38.110903683955435, 9.899495124816895, 294.7808837890625, 292.3835754394531, 293.81494140625, 291.11065673828125, 293.51739501953125, 283.6424865722656, 13.580912590026855, 4.976086616516113, 1.1271398067474365, 0.9465181231498718, 0.5066819190979004, 0.1808401197195053, 0.0]
17463699
[0.0, 0.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0]
def overall_average_score(actual,prediction):
precision = precision_recall_fscore_support(actual, prediction, average = 'binary')[0]
recall = precision_recall_fscore_support(actual, prediction, average = 'binary')[1]
f1_score = precision_recall_fscore_support(actual, prediction, average = 'binary')[2]
total_score = matthews_corrcoef(actual, prediction)+accuracy_score(actual, prediction)+precision+recall+f1_score
return total_score/5
grid_scorer = make_scorer(overall_average_score, greater_is_better=True)
parameters = {'n_estimators': [10,20,30], 'max_features': ['auto','sqrt','log2',0.5,0.3], }
random = RandomForestClassifier()
clf = grid_search.GridSearchCV(random, parameters, cv = 5, scoring = grid_scorer)
clf.fit(X,Y)
ValueError Traceback (most recent call last)
<ipython-input-39-a8686eb798b2> in <module>()
18 random = RandomForestClassifier()
19 clf = grid_search.GridSearchCV(random, parameters, cv = 5, scoring = grid_scorer)
---> 20 clf.fit(X,Y)
21
22
/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/grid_search.pyc in fit(self, X, y)
730
731 """
--> 732 return self._fit(X, y, ParameterGrid(self.param_grid))
733
734
/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/grid_search.pyc in _fit(self, X, y, parameter_iterable)
503 self.fit_params, return_parameters=True,
504 error_score=self.error_score)
--> 505 for parameters in parameter_iterable
506 for train, test in cv)
507
/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/externals/joblib/parallel.pyc in __call__(self, iterable)
657 self._iterating = True
658 for function, args, kwargs in iterable:
--> 659 self.dispatch(function, args, kwargs)
660
661 if pre_dispatch == "all" or n_jobs == 1:
/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/externals/joblib/parallel.pyc in dispatch(self, func, args, kwargs)
404 """
405 if self._pool is None:
--> 406 job = ImmediateApply(func, args, kwargs)
407 index = len(self._jobs)
408 if not _verbosity_filter(index, self.verbose):
/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/externals/joblib/parallel.pyc in __init__(self, func, args, kwargs)
138 # Don't delay the application, to avoid keeping the input
139 # arguments in memory
--> 140 self.results = func(*args, **kwargs)
141
142 def get(self):
/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/cross_validation.pyc in _fit_and_score(estimator, X, y, scorer, train, test, verbose, parameters, fit_params, return_train_score, return_parameters, error_score)
1476
1477 else:
-> 1478 test_score = _score(estimator, X_test, y_test, scorer)
1479 if return_train_score:
1480 train_score = _score(estimator, X_train, y_train, scorer)
/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/cross_validation.pyc in _score(estimator, X_test, y_test, scorer)
1532 score = scorer(estimator, X_test)
1533 else:
-> 1534 score = scorer(estimator, X_test, y_test)
1535 if not isinstance(score, numbers.Number):
1536 raise ValueError("scoring must return a number, got %s (%s) instead."
/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/metrics/scorer.pyc in __call__(self, estimator, X, y_true, sample_weight)
87 else:
88 return self._sign * self._score_func(y_true, y_pred,
---> 89 **self._kwargs)
90
91
<ipython-input-39-a8686eb798b2> in overall_average_score(actual, prediction)
3 recall = precision_recall_fscore_support(actual, prediction, average = 'binary')[1]
4 f1_score = precision_recall_fscore_support(actual, prediction, average = 'binary')[2]
----> 5 total_score = matthews_corrcoef(actual, prediction)+accuracy_score(actual, prediction)+precision+recall+f1_score
6 return total_score/5
7 def show_score(actual,prediction):
/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/metrics/classification.pyc in matthews_corrcoef(y_true, y_pred)
395
396 if y_type != "binary":
--> 397 raise ValueError("%s is not supported" % y_type)
398
399 lb = LabelEncoder()
ValueError: multiclass is not supported
马修斯相关系数是介于-1到1之间的分数。因此，计算f1_score、precision、recall、accuracy_score和MCC之间的平均值是不正确的。
MCC值表明:1为总正相关0表示不相关−1为总负相关
而上面提到的其他评价指标在0到1之间(从最差到最好的精度指标)。
我复制了你的实验，但我没有得到任何错误。此错误指示向量actual或prediction 包含两个以上的离散值
你能够在GridSearchCV之外训练的随机森林得分，这确实很奇怪。
您能提供执行此操作的确切代码吗?
下面是我用来重现错误的代码:
from sklearn.datasets import make_classification
from sklearn.grid_search import GridSearchCV
from sklearn.metrics import precision_recall_fscore_support, accuracy_score,
matthews_corrcoef, make_scorer
from sklearn.ensemble import RandomForestClassifier
from sklearn.cross_validation import train_test_split
def overall_average_score(actual,prediction):
precision, recall, f1_score, _ = precision_recall_fscore_support(
actual, prediction, average='binary')
total_score = (matthews_corrcoef(actual, prediction) +
accuracy_score(actual, prediction) + precision + recall + f1_score)
return total_score / 5
grid_scorer = make_scorer(overall_average_score, greater_is_better=True)
print("Without GridSearchCV")
X, y = make_classification(n_samples=500, n_informative=10, n_classes=2)
X_train, X_test, y_train, y_test = train_test_split(X, y,
test_size=0.5, random_state=0)
rf = RandomForestClassifier()
rf.fit(X_train, y_train)
y_pred = rf.predict(X_test)
print("Overall average score: ", overall_average_score(y_test, y_pred))
print("-" * 30)
print("With GridSearchCV:")
parameters = {'n_estimators': [10,20,30],
'max_features': ['auto','sqrt','log2',0.5,0.3], }
gs_rf = GridSearchCV(rf, parameters, cv=5, scoring=grid_scorer)
gs_rf.fit(X_train,y_train)
print("Best score with grid search: ", gs_rf.best_score_)
现在我想对你提供的代码做一些注释:
- 使用诸如
- 你可以直接解包
precision,recall和f1_score，而不是调用3次precision_recall_fscore_support。 - 在
n_estimators上进行网格搜索并没有真正意义:更多的树总是更好的。如果您担心过拟合，您可以通过使用其他参数(如max_depth或min_samples_split)来降低单个模型的复杂性。
random(这通常是一个模块)或f1_score(这与sklearn.metrics.f1_score方法冲突)之类的变量名并不是一个很好的实践。