使用正则化 (L1 / L2) Lasso 和 Ridge 的 Logistic 回归模型

Question

我正在尝试构建模型并创建网格搜索，下面是代码。原始数据是从该网站下载的（信用卡欺诈数据）。 https://www.kaggle.com/mlg-ulb/creditcardfraud读取数据后从标准化开始编码。standardization = StandardScaler()credit_card_fraud_df[['Amount']] = standardization.fit_transform(credit_card_fraud_df[['Amount']])# Assigning feature variable to XX = credit_card_fraud_df.drop(['Class'], axis=1)# Assigning response variable to yy = credit_card_fraud_df['Class']# Splitting the data into train and testX_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.7, test_size=0.3, random_state=100)X_train.head()power_transformer = PowerTransformer(copy=False)power_transformer.fit(X_train)                       ## Fit the PT on training dataX_train_pt_df = power_transformer.transform(X_train)    ## Then apply on all dataX_test_pt_df = power_transformer.transform(X_test)y_train_pt_df = y_trainy_test_pt_df = y_testtrain_pt_df = pd.DataFrame(data=X_train_pt_df, columns=X_train.columns.tolist())# set up cross validation schemefolds = StratifiedKFold(n_splits = 5, shuffle = True, random_state = 4)# specify range of hyperparametersparams = {"C":np.logspace(-3,3,5,7), "penalty":["l1","l2"]}# l1 lasso l2 ridge结果示例：  mean_fit_time std_fit_time    mean_score_time std_score_time  param_C param_penalty   params  split0_test_score   split1_test_score   split2_test_score   split3_test_score   split4_test_score   mean_test_score std_test_score  rank_test_score    0   0.044332    0.002040    0.000000    0.000000    0.001   l1  {'C': 0.001, 'penalty': 'l1'}   NaN NaN NaN NaN NaN NaN NaN 6    1   0.477965    0.046651    0.016745    0.003813    0.001   l2  {'C': 0.001, 'penalty': 'l2'}   0.485714    0.428571    0.542857    0.485714    0.457143    0.480000    0.037904    5我的输入数据中没有任何空值。我不明白为什么我会得到这些列的 Nan 值。谁能帮帮我吗？

Answer 1

您在此处定义的默认求解器有问题：

model = LogisticRegression(class_weight='balanced')

这是从以下错误消息得出的：

ValueError: Solver lbfgs supports only 'l2' or 'none' penalties, got l1 penalty.

此外，在定义参数网格之前研究文档可能会很有用：

penalty: {'l1', 'l2', 'elasticnet', 'none'}, default='l2' 用于指定惩罚中使用的范数。“newton-cg”、“sag”和“lbfgs”求解器仅支持 l2 惩罚。“elasticnet”仅受“saga”求解器支持。如果为“none”（liblinear 求解器不支持），则不应用正则化。

一旦您使用支持所需网格的不同解算器纠正它，您就可以开始：

## using Logistic regression for class imbalance

model = LogisticRegression(class_weight='balanced', solver='saga')

grid_search_cv = GridSearchCV(estimator = model, param_grid = params, 

                        scoring= 'roc_auc', 

                        cv = folds, 

                        return_train_score=True, verbose = 1)            

grid_search_cv.fit(X_train_pt_df, y_train_pt_df)

## reviewing the results

cv_results = pd.DataFrame(grid_search_cv.cv_results_)

另请注意，ConvergenceWarning这可能建议您需要增加默认值max_iter、tol或切换到另一个求解器并重新考虑所需的参数网格。