使用LR和XGBoost跑通criteo点击率预测数据集
2024-05-10 10:27:52
SuperCTR
评估指标
from sklearn.metrics import accuracy_score
from sklearn.metrics import precision_score
from sklearn.metrics import recall_score
from sklearn.metrics import f1_score
from sklearn.metrics import roc_auc_score
def eval_model(y_pred, y_true):print(f"accuracy_score = {accuracy_score(y_true, y_pred)}")print(f"precision_score = {precision_score(y_true, y_pred)}")print(f"recall_score = {recall_score(y_true, y_pred)}")print(f"f1_score = {f1_score(y_true, y_pred)}")print(f"auc = {roc_auc_score(y_true, y_pred)}")
LR
1.导包
import time
import numpy as np
import pandas as pd
import seaborn as sns
import xgboost as xgb
import matplotlib.pyplot as plt
from xgboost import XGBClassifier
from sklearn.model_selection import GridSearchCV
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.model_selection import KFold
2.导入数据
train_df = pd.read_csv("../input/criteo-dataset/dac/train.txt", sep='\t', names=train_df_cols, nrows=1000000)
train_df.head()
test_df = pd.read_csv("../input/criteo-dataset/dac/test.txt", sep='\t', names=test_df_cols)
数据共40个特征，其中特征连续型的有13个（I1-I13），类别型的26个（C1-C26），一个标签（Label）。
3.对类别型的数据编码
因为只有编码之后才能用于模型的训练
continuous_variable = [f"I{i}" for i in range(1, 14)]
discrete_variable = [f"C{i}" for i in range(1, 27)]
train_df_cols = ["Label"] + continuous_variable + discrete_variable
test_df_cols = continuous_variable + discrete_variable
print(f"train_df has {train_df.shape[0]} rows and {train_df.shape[1]} columns.")
print(f"test_df has {test_df.shape[0]} rows and {test_df.shape[1]} columns.")
将类别型的每一列用factorize方法进行编码，并用均值填补空缺值
for col in train_df.columns:if train_df[col].dtypes == "object":train_df[col], uniques = pd.factorize(train_df[col])train_df[col].fillna(train_df[col].mean(), inplace=True)
train_df.head()
train_df.info()
结果如下，没有空值了。
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 1000000 entries, 0 to 999999
Data columns (total 40 columns):# Column Non-Null Count Dtype
--- ------ -------------- ----- 0 Label 1000000 non-null int64 1 I1 1000000 non-null float642 I2 1000000 non-null int64 3 I3 1000000 non-null float644 I4 1000000 non-null float645 I5 1000000 non-null float646 I6 1000000 non-null float647 I7 1000000 non-null float648 I8 1000000 non-null float649 I9 1000000 non-null float6410 I10 1000000 non-null float6411 I11 1000000 non-null float6412 I12 1000000 non-null float6413 I13 1000000 non-null float6414 C1 1000000 non-null int64 15 C2 1000000 non-null int64 16 C3 1000000 non-null int64 17 C4 1000000 non-null int64 18 C5 1000000 non-null int64 19 C6 1000000 non-null int64 20 C7 1000000 non-null int64 21 C8 1000000 non-null int64 22 C9 1000000 non-null int64 23 C10 1000000 non-null int64 24 C11 1000000 non-null int64 25 C12 1000000 non-null int64 26 C13 1000000 non-null int64 27 C14 1000000 non-null int64 28 C15 1000000 non-null int64 29 C16 1000000 non-null int64 30 C17 1000000 non-null int64 31 C18 1000000 non-null int64 32 C19 1000000 non-null int64 33 C20 1000000 non-null int64 34 C21 1000000 non-null int64 35 C22 1000000 non-null int64 36 C23 1000000 non-null int64 37 C24 1000000 non-null int64 38 C25 1000000 non-null int64 39 C26 1000000 non-null int64
dtypes: float64(12), int64(28)
memory usage: 305.2 MB
画一下标签的频率直方图（0表示未点击，1表示点击）
count_label = pd.value_counts(train_df["Label"], sort=True)
count_label.plot(kind="bar") # 条形图
plt.title("Label Statistics")
plt.xlabel("Label")
plt.ylabel("Frequency")
算一下具体的正负样本比例
#负样本:正样本 = 2.92:1（scale_pos_weight=2.92）
pd.value_counts(train_df["Label"], sort=True)
下采样
# 下采样
number_of_click = len(train_df[train_df.Label == 1]) # 统计点击的数据量
click_indices = np.array(train_df[train_df.Label == 1].index) # 点击的数据索引
no_click_indices = np.array(train_df[train_df.Label == 0].index) # 正常数据索引# 在正常数据中随机选择与未点击数据量相等的正常数据的索引
random_no_click_indices = np.array(np.random.choice(no_click_indices, number_of_click, replace=False))# 将所有点击的数据索引和未点击的等量数据索引合并
under_sample_indices = np.concatenate([click_indices, random_no_click_indices])
# 从原始数据中取出下采样数据集
under_sample_train_df = train_df.iloc[under_sample_indices, :]X_under_sample = under_sample_train_df.iloc[:, under_sample_train_df.columns != "Label"]
Y_under_sample = under_sample_train_df.iloc[:, under_sample_train_df.columns == "Label"]
# 看一下下采样数据集的长度
print("Total number of under_sample_train_df =", len(under_sample_train_df))
print("Total number of no_click =", len(under_sample_train_df[train_df.Label == 0]))
print("Total number of click =", len(under_sample_train_df[train_df.Label == 1]))
结果如下，正负样本比例已经均衡。
Total number of under_sample_train_df = 509898
Total number of no_click = 254949
Total number of click = 254949
X_under_sample.head()
Y_under_sample
X_under_sample_train,X_under_sample_test,y_under_sample_train,y_under_sample_test = train_test_split(X_under_sample,Y_under_sample,test_size=0.3)
# 对下采样数据进行测试
lr = LogisticRegression(solver="liblinear")
lr.fit(X_under_sample, Y_under_sample)
lr.score(X_under_sample_test, y_under_sample_test)
得到结果：0.6450088252598549
看一下各项评价指标
y_predict_lr = lr.predict(X_under_sample_test)
eval_model(y_predict_lr, y_under_sample_test)
其实效果并不算好，总体才60%多。接下来试试xgboost跑出来的效果。
xgboost
做k折交叉验证
# 下采样后做交叉验证
def kfold_scores(x_train_data, y_train_data):start_time = time.time()fold = KFold(3, shuffle=True) # 3折交叉验证#c_param_range = [0, 0.5, 1] # 惩罚力度，正则化惩罚项的系数# 做可视化展示results_table = pd.DataFrame(index=range(2),columns=["C_parameter", "Mean recall scores", "Mean auc socres"])index = 0print('--------------------------------------------------------------------------------')# 做交叉验证recall_accs = []auc_scores = []# 可以加入scale_pos_weight=2.92参数，如果使用下采样就不加入scale_pos_weight以免互相影响xgb_model = XGBClassifier(objective="binary:logistic", n_jobs=-1, n_estimators=1000, max_depth=16,eval_metric="auc",colsample_bytree=0.8, subsample=0.8, learning_rate=0.2, min_child_weight=6)for iteration, indices in enumerate(fold.split(x_train_data)):# 拟合训练数据# lr.fit(x_train_data.iloc[indices[0], :], y_train_data.iloc[indices[0], :].values.ravel())xgb_model.fit(x_train_data.iloc[indices[0], :], y_train_data.iloc[indices[0], :].values.ravel(),eval_metric="logloss",eval_set=[(x_train_data.iloc[indices[0], :], y_train_data.iloc[indices[0], :]),(x_train_data.iloc[indices[1], :], y_train_data.iloc[indices[1], :])], verbose=True,early_stopping_rounds=10)# 使用验证集得出预测数据# 最适合的迭代次数，然后预测的时候就使用stop之前训练的树来预测。print(f"最佳迭代次数为：{xgb_model.best_iteration}")limit = xgb_model.best_iteration# y_predicted_undersample = lr.predict(x_train_data.iloc[indices[1], :])y_predicted_undersample = xgb_model.predict(x_train_data.iloc[indices[1], :], ntree_limit=limit)# 计算recallrecall_acc = recall_score(y_train_data.iloc[indices[1], :], y_predicted_undersample)recall_accs.append(recall_acc)auc_score = roc_auc_score(y_train_data.iloc[indices[1], :], y_predicted_undersample)auc_scores.append(auc_score)print('\tIteration ', iteration, ': recall score = ', recall_acc)print('\tIteration ', iteration, ': auc score = ', auc_score)index += 1# 计算recall的平均值results_table.loc[index, "Mean recall scores"] = np.mean(recall_accs)results_table.loc[index, "Mean auc scores"] = np.mean(auc_scores)print('Mean recall score = ', results_table.loc[index, "Mean recall scores"], end="\n\n")print('Mean auc score = ', results_table.loc[index, "Mean auc scores"], end="\n\n")print('--------------------------------------------------------------------------------')return xgb_model
# 对下采样数据进行测试
xgb = kfold_scores(X_under_sample, Y_under_sample)
训练结果如下
--------------------------------------------------------------------------------
[0] validation_0-logloss:0.65029 validation_1-logloss:0.66628
[1] validation_0-logloss:0.61773 validation_1-logloss:0.64819
[2] validation_0-logloss:0.59209 validation_1-logloss:0.63575
[3] validation_0-logloss:0.57027 validation_1-logloss:0.62659
[4] validation_0-logloss:0.55167 validation_1-logloss:0.62060
[5] validation_0-logloss:0.53649 validation_1-logloss:0.61567
[6] validation_0-logloss:0.52190 validation_1-logloss:0.61203
[7] validation_0-logloss:0.50868 validation_1-logloss:0.60933
[8] validation_0-logloss:0.49658 validation_1-logloss:0.60661
[9] validation_0-logloss:0.48651 validation_1-logloss:0.60501
[10] validation_0-logloss:0.47802 validation_1-logloss:0.60350
[11] validation_0-logloss:0.46839 validation_1-logloss:0.60220
[12] validation_0-logloss:0.46043 validation_1-logloss:0.60108
[13] validation_0-logloss:0.45288 validation_1-logloss:0.60091
[14] validation_0-logloss:0.44480 validation_1-logloss:0.60020
[15] validation_0-logloss:0.43958 validation_1-logloss:0.60004
[16] validation_0-logloss:0.43263 validation_1-logloss:0.59916
[17] validation_0-logloss:0.42796 validation_1-logloss:0.59871
[18] validation_0-logloss:0.42143 validation_1-logloss:0.59718
[19] validation_0-logloss:0.41582 validation_1-logloss:0.59657
[20] validation_0-logloss:0.41130 validation_1-logloss:0.59631
[21] validation_0-logloss:0.40725 validation_1-logloss:0.59608
[22] validation_0-logloss:0.40239 validation_1-logloss:0.59571
[23] validation_0-logloss:0.39832 validation_1-logloss:0.59539
[24] validation_0-logloss:0.39447 validation_1-logloss:0.59505
[25] validation_0-logloss:0.39110 validation_1-logloss:0.59487
[26] validation_0-logloss:0.38895 validation_1-logloss:0.59436
[27] validation_0-logloss:0.38441 validation_1-logloss:0.59404
[28] validation_0-logloss:0.38127 validation_1-logloss:0.59416
[29] validation_0-logloss:0.37733 validation_1-logloss:0.59403
[30] validation_0-logloss:0.37380 validation_1-logloss:0.59427
[31] validation_0-logloss:0.37197 validation_1-logloss:0.59412
[32] validation_0-logloss:0.36929 validation_1-logloss:0.59381
[33] validation_0-logloss:0.36792 validation_1-logloss:0.59378
[34] validation_0-logloss:0.36600 validation_1-logloss:0.59374
[35] validation_0-logloss:0.36452 validation_1-logloss:0.59383
[36] validation_0-logloss:0.36155 validation_1-logloss:0.59420
[37] validation_0-logloss:0.35944 validation_1-logloss:0.59413
[38] validation_0-logloss:0.35899 validation_1-logloss:0.59409
[39] validation_0-logloss:0.35702 validation_1-logloss:0.59423
[40] validation_0-logloss:0.35486 validation_1-logloss:0.59396
[41] validation_0-logloss:0.35240 validation_1-logloss:0.59411
[42] validation_0-logloss:0.35000 validation_1-logloss:0.59412
[43] validation_0-logloss:0.34818 validation_1-logloss:0.59408
最佳迭代次数为：34
/opt/conda/lib/python3.7/site-packages/xgboost/core.py:104: UserWarning: ntree_limit is deprecated, use `iteration_range` or model slicing instead.UserWarningIteration 0 : recall score = 0.6809424858239707Iteration 0 : auc score = 0.6823219983343968
/opt/conda/lib/python3.7/site-packages/xgboost/sklearn.py:1146: UserWarning: The use of label encoder in XGBClassifier is deprecated and will be removed in a future release. To remove this warning, do the following: 1) Pass option use_label_encoder=False when constructing XGBClassifier object; and 2) Encode your labels (y) as integers starting with 0, i.e. 0, 1, 2, ..., [num_class - 1].warnings.warn(label_encoder_deprecation_msg, UserWarning)
/opt/conda/lib/python3.7/site-packages/sklearn/utils/validation.py:72: DataConversionWarning: A column-vector y was passed when a 1d array was expected. Please change the shape of y to (n_samples, ), for example using ravel().return f(**kwargs)
[0] validation_0-logloss:0.65061 validation_1-logloss:0.66599
[1] validation_0-logloss:0.61815 validation_1-logloss:0.64769
[2] validation_0-logloss:0.59309 validation_1-logloss:0.63531
[3] validation_0-logloss:0.57095 validation_1-logloss:0.62661
[4] validation_0-logloss:0.55278 validation_1-logloss:0.62015
[5] validation_0-logloss:0.53753 validation_1-logloss:0.61507
[6] validation_0-logloss:0.52324 validation_1-logloss:0.61146
[7] validation_0-logloss:0.51051 validation_1-logloss:0.60895
[8] validation_0-logloss:0.49835 validation_1-logloss:0.60653
[9] validation_0-logloss:0.48860 validation_1-logloss:0.60473
[10] validation_0-logloss:0.47981 validation_1-logloss:0.60401
[11] validation_0-logloss:0.47103 validation_1-logloss:0.60256
[12] validation_0-logloss:0.46264 validation_1-logloss:0.60144
[13] validation_0-logloss:0.45640 validation_1-logloss:0.60092
[14] validation_0-logloss:0.44830 validation_1-logloss:0.60041
[15] validation_0-logloss:0.44047 validation_1-logloss:0.59898
[16] validation_0-logloss:0.43539 validation_1-logloss:0.59905
[17] validation_0-logloss:0.43033 validation_1-logloss:0.59883
[18] validation_0-logloss:0.42509 validation_1-logloss:0.59805
[19] validation_0-logloss:0.41953 validation_1-logloss:0.59732
[20] validation_0-logloss:0.41452 validation_1-logloss:0.59687
[21] validation_0-logloss:0.40779 validation_1-logloss:0.59526
[22] validation_0-logloss:0.40325 validation_1-logloss:0.59515
[23] validation_0-logloss:0.39959 validation_1-logloss:0.59504
[24] validation_0-logloss:0.39437 validation_1-logloss:0.59415
[25] validation_0-logloss:0.39129 validation_1-logloss:0.59403
[26] validation_0-logloss:0.38813 validation_1-logloss:0.59418
[27] validation_0-logloss:0.38452 validation_1-logloss:0.59421
[28] validation_0-logloss:0.38126 validation_1-logloss:0.59379
[29] validation_0-logloss:0.37753 validation_1-logloss:0.59386
[30] validation_0-logloss:0.37584 validation_1-logloss:0.59387
[31] validation_0-logloss:0.37175 validation_1-logloss:0.59363
[32] validation_0-logloss:0.36911 validation_1-logloss:0.59344
[33] validation_0-logloss:0.36628 validation_1-logloss:0.59371
[34] validation_0-logloss:0.36368 validation_1-logloss:0.59380
[35] validation_0-logloss:0.36148 validation_1-logloss:0.59360
[36] validation_0-logloss:0.35812 validation_1-logloss:0.59368
[37] validation_0-logloss:0.35587 validation_1-logloss:0.59382
[38] validation_0-logloss:0.35382 validation_1-logloss:0.59387
[39] validation_0-logloss:0.35260 validation_1-logloss:0.59375
[40] validation_0-logloss:0.35016 validation_1-logloss:0.59354
[41] validation_0-logloss:0.34815 validation_1-logloss:0.59372
最佳迭代次数为：32
/opt/conda/lib/python3.7/site-packages/xgboost/core.py:104: UserWarning: ntree_limit is deprecated, use `iteration_range` or model slicing instead.UserWarningIteration 1 : recall score = 0.6808235901770202Iteration 1 : auc score = 0.6835137790937945
/opt/conda/lib/python3.7/site-packages/xgboost/sklearn.py:1146: UserWarning: The use of label encoder in XGBClassifier is deprecated and will be removed in a future release. To remove this warning, do the following: 1) Pass option use_label_encoder=False when constructing XGBClassifier object; and 2) Encode your labels (y) as integers starting with 0, i.e. 0, 1, 2, ..., [num_class - 1].warnings.warn(label_encoder_deprecation_msg, UserWarning)
/opt/conda/lib/python3.7/site-packages/sklearn/utils/validation.py:72: DataConversionWarning: A column-vector y was passed when a 1d array was expected. Please change the shape of y to (n_samples, ), for example using ravel().return f(**kwargs)
[0] validation_0-logloss:0.65044 validation_1-logloss:0.66610
[1] validation_0-logloss:0.61792 validation_1-logloss:0.64771
[2] validation_0-logloss:0.59264 validation_1-logloss:0.63498
[3] validation_0-logloss:0.57097 validation_1-logloss:0.62585
[4] validation_0-logloss:0.55261 validation_1-logloss:0.61955
[5] validation_0-logloss:0.53678 validation_1-logloss:0.61493
[6] validation_0-logloss:0.52250 validation_1-logloss:0.61141
[7] validation_0-logloss:0.50931 validation_1-logloss:0.60843
[8] validation_0-logloss:0.49780 validation_1-logloss:0.60593
[9] validation_0-logloss:0.48695 validation_1-logloss:0.60404
[10] validation_0-logloss:0.47770 validation_1-logloss:0.60325
[11] validation_0-logloss:0.46977 validation_1-logloss:0.60226
[12] validation_0-logloss:0.46171 validation_1-logloss:0.60099
[13] validation_0-logloss:0.45432 validation_1-logloss:0.60052
[14] validation_0-logloss:0.44628 validation_1-logloss:0.59974
[15] validation_0-logloss:0.43920 validation_1-logloss:0.59892
[16] validation_0-logloss:0.43217 validation_1-logloss:0.59798
[17] validation_0-logloss:0.42623 validation_1-logloss:0.59718
[18] validation_0-logloss:0.41996 validation_1-logloss:0.59606
[19] validation_0-logloss:0.41392 validation_1-logloss:0.59542
[20] validation_0-logloss:0.40929 validation_1-logloss:0.59504
[21] validation_0-logloss:0.40538 validation_1-logloss:0.59497
[22] validation_0-logloss:0.40013 validation_1-logloss:0.59454
[23] validation_0-logloss:0.39680 validation_1-logloss:0.59428
[24] validation_0-logloss:0.39408 validation_1-logloss:0.59417
[25] validation_0-logloss:0.39081 validation_1-logloss:0.59376
[26] validation_0-logloss:0.38750 validation_1-logloss:0.59376
[27] validation_0-logloss:0.38438 validation_1-logloss:0.59282
[28] validation_0-logloss:0.38091 validation_1-logloss:0.59292
[29] validation_0-logloss:0.37594 validation_1-logloss:0.59296
[30] validation_0-logloss:0.37254 validation_1-logloss:0.59232
[31] validation_0-logloss:0.36862 validation_1-logloss:0.59233
[32] validation_0-logloss:0.36653 validation_1-logloss:0.59235
[33] validation_0-logloss:0.36366 validation_1-logloss:0.59235
进行预测
y_predict_xgb = xgb.predict(X_under_sample_test,ntree_limit=xgb.best_iteration)
eval_model(y_predict_xgb, y_under_sample_test)
即
accuracy_score = 0.8150748512780284
precision_score = 0.8165117163425604
recall_score = 0.8132910152423495
f1_score = 0.8148981835313827
auc = 0.8150766723050388
嗯，好，这效果就非常地、大大地好！
n_0-logloss:0.36366 validation_1-logloss:0.59235
进行预测
y_predict_xgb = xgb.predict(X_under_sample_test,ntree_limit=xgb.best_iteration)
eval_model(y_predict_xgb, y_under_sample_test)
即
accuracy_score = 0.8150748512780284
precision_score = 0.8165117163425604
recall_score = 0.8132910152423495
f1_score = 0.8148981835313827
auc = 0.8150766723050388
嗯，好，这效果就非常地、大大地好！
使用LR和XGBoost跑通criteo点击率预测数据集相关推荐
- criteo 点击率预估_预处理criteo数据集以预测广告的点击率
criteo 点击率预估 Amany Abdelhalim阿曼尼·阿卜杜勒哈林 Follow跟随 Sep 18 九月18 Preprocessing Criteo Dataset for Predic ...
- 使用GBDT+LR作点击率预测
主要内容来源于facebook的论文:Practical Lessons from Predicting Clicks on Ads at Facebook> 1.基本思路 使用GBDT根据用户 ...
- 点云3D目标检测之——尝试SFD代码跑通（超详细！！）
前言 到目前为止还没跑通,但是bug实在太多了,我的每一步都有错,如果不记录下来又会有遗漏,(肯定已经遗漏了很多),在这里把能想起来的都记录一下以便不时之需.另外,本人深度学习小白,一上来跑这么难的代 ...
- 跑通Yolov5的一些心得
跑通Yolov5的一些心得 1.文件安装准备(版本要求) #pytorch 1.5.1 #torchvision 0.6.1 #vs2015 conda create -n yolov5 python ...
- LeNet-5N 网络模型及对应代码自用已跑通
先来结构图 再上数据结果 # 三个列表对应代码中的三个print的结果 # 这个是epoch的loss 波动很大,所以再绘制时我用的另一个列表为train_loss_sum_batch_size_li ...
- 深度学习之初识篇——小白也能跑通的深度学习万能框架【交通标识牌检测】
目录 环境下载:点击即可 数据集下载:点击即可 深度学习环境配置 点击下载深度学习环境 数据集准备 使用自己标注的数据集 使用标注软件 数据准备 VOC标签格式转yolo格式并划分训练集和测试集 部署 ...
- 【全网唯一】全网唯一能够跑通的，跑不通你来找我~用node.js完成微信支付下单功能，且只需要一个文件wxpay.js就解决业务流程的node.js程序
先吐为敬! 最近心血来潮研究nodejs如何完成微信支付功能,结果网上一搜索,一大堆"代码拷贝党"."留一手"."缺斤少两"."不 ...
- 13个月才跑通GitHub模型，机器学习科研入门太难了吧
点击上方"视学算法",选择加"星标"或"置顶" 重磅干货,第一时间送达 金磊 萧箫 发自 凹非寺 量子位 报道 | 公众号 QbitAI & ...
- @卡尔曼滤波 跑通调参
可以先像一开始接触PID那样,先把整个系统跑通,感受一下它是咋用的.后面再回过头来研究原理. 学习OpenCV2--卡尔曼滤波(KalmanFilter)详解 四旋翼姿态解算--互补滤波和拓展卡尔曼 ...
最新文章
- 详解 Python 如何将爬取到的数据分别存储到 txt、excel、mysql 中！
- 微生物组学研究的那些”奇葩“动物模型
- mysql服务器的启动方式有哪几种_Mysql启动的方式(四种)
- ironbot智能编程机器人_视频 | 多模式编程机器人，“程序猿”培养从小抓起
- 深入理解maven及应用--转
- 数据科学-通过数据探索了解我们的特征
- IAR无法goto的解决办法
- UFLDL教程： Exercise: Sparse Autoencoder
- python判断互质_整数判断是否互质并求逆元的 python 实现
- jQuery判断页面是电脑端还是手机端
- wsgiserver python 漏洞_python-简单测试wsgi
- Net设计模式实例之桥接模式( Bridge Pattern)(1)
- java snmpv3_snmpv3 java实现
- eclipse安装SVN插件（2020最新，亲测可用）
- 网易有道最新力作 有道词典笔3 结构拆解
- 基于W800的AIOT离在线一体方案说明 （阿里飞燕+离线语音控制）
- selenium鼠标操作
- VMware虚拟机Mac-OS-X系统如何切换中英文输入法
- C# 串口接收含有asic2=0x3f时发生的故障
- 第一次用idea写SSM框架的项目就遇到自动注入失败的问题