为 IsolationForest GridSearchCV 创建评分指标

Question

我正在做一个 IsolationForest，我想使用 GridSearchCV 优化我的超参数。我希望我的评分基于异常值的召回分数，即标签 = -1。但是，我在运行此代码时遇到了错误。

recall_fraud = make_scorer(recall_score(pos_label=-1))
gs_params ={
        'max_samples': [300,500,1000],
        'contamination': [float(y_train.count(-1))/len(y_train)] ,
        'max_features': [1,3,7],
        'n_estimators':[1000],
        'random_state':[1]
    }

isof_gs = GridSearchCV(IsolationForest(), gs_params, n_jobs = 1, verbose = 1, cv = 5, scoring = recall_fraud) 

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-273-d1f260f73f29> in <module>()
----> 1 recall_fraud = make_scorer(recall_score(pos_label=-1))
      2 gs_params ={
      3     'max_samples': [300,500,1000],
      4     'contamination': [float(y_train.count(-1))/len(y_train)] ,
      5     'max_features': [1,3,7],

TypeError: recall_score() takes at least 2 arguments (1 given)

我做错了什么吗？

Answer 1

使用make_scorer 时，您要传递给评分函数的所有关键字参数都应传递给make_scorer，而不是内部评分函数。

看kwargs param in make_scorer:-

**kwargs : 附加参数 要传递给 score_func 的附加参数。

像这样更改您的代码：

# Updated recall_score() to recall_score
recall_fraud = make_scorer(recall_score, pos_label=-1)

它不会再抛出错误了。

pos_label在实际计算grid-search的recall时会自动转发到recall_score方法。