2022讯飞——糖尿病遗传风险检测挑战赛解决方案

0. 赛事背景

截至2022年，中国糖尿病患者近1.3亿。中国糖尿病患病原因受生活方式、老龄化、城市化、家族遗传等多种因素影响。同时，糖尿病患者趋向年轻化。

糖尿病可导致心血管、肾脏、脑血管并发症的发生。因此，准确诊断出患有糖尿病个体具有非常重要的临床意义。糖尿病早期遗传风险预测将有助于预防糖尿病的发生。

根据《中国2型糖尿病防治指南（2017年版）》，糖尿病的诊断标准是具有典型糖尿病症状（烦渴多饮、多尿、多食、不明原因的体重下降）且随机静脉血浆葡萄糖≥11.1mmol/L或空腹静脉血浆葡萄糖≥7.0mmol/L或口服葡萄糖耐量试验（OGTT）负荷后2h血浆葡萄糖≥11.1mmol/L。

在这次比赛中，您需要通过训练数据集构建糖尿病遗传风险预测模型，然后预测出测试数据集中个体是否患有糖尿病，和我们一起帮助糖尿病患者解决这“甜蜜的烦恼”。

• 训练集说明
  训练集（比赛训练集.csv）一共有5070条数据，用于构建您的预测模型（您可能需要先进行数据分析）。数据的字段有编号、性别、出生年份、体重指数、糖尿病家族史、舒张压、口服耐糖量测试、胰岛素释放实验、肱三头肌皮褶厚度、患有糖尿病标识（最后一列），您也可以通过特征工程技术构建新的特征。

• 测试集说明
  测试集（比赛测试集.csv）一共有1000条数据，用于验证预测模型的性能。数据的字段有编号、性别、出生年份、体重指数、糖尿病家族史、舒张压、口服耐糖量测试、胰岛素释放实验、肱三头肌皮褶厚度。

• 评估指标
  对于提交的结果，系统会采用二分类任务中的F1-score指标进行评价，F1-score越大说明预测模型性能越好。

本质为二分类预测，不是很难，注意做好特征工程，提升数据质量。

比赛详情见官网。

1. 读取数据

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import lightgbm

train_data = pd.read_csv('比赛训练集.csv',encoding='gbk')
test_data = pd.read_csv('比赛测试集.csv',encoding='gbk')
train_data.describe()

	编号	性别	出生年份	体重指数	舒张压	口服耐糖量测试	胰岛素释放实验	肱三头肌皮褶厚度	患有糖尿病标识
count	5070.000000	5070.000000	5070.000000	5070.000000	4823.000000	5070.000000	5070.000000	5070.000000	5070.000000
mean	2535.500000	0.456805	1986.869231	37.986785	89.423595	5.612839	4.114321	6.994371	0.381854
std	1463.727263	0.498180	8.919737	11.447095	9.266992	2.257649	8.726001	13.651442	0.485889
min	1.000000	0.000000	1943.000000	0.000000	30.000000	-1.000000	0.000000	0.000000	0.000000
25%	1268.250000	0.000000	1980.000000	28.400000	85.000000	4.314000	0.000000	0.000000	0.000000
50%	2535.500000	0.000000	1987.000000	36.550000	89.000000	5.760000	0.000000	0.000000	0.000000
75%	3802.750000	1.000000	1995.000000	47.600000	96.000000	7.193000	7.100000	4.120000	1.000000
max	5070.000000	1.000000	2009.000000	65.900000	126.000000	10.839000	108.960000	45.000000	1.000000

患有糖尿病标识的均值为0.38，说明未患病：患病 $\approx 6:4$，数据还算较为平衡。

test_data.describe()

	编号	性别	出生年份	体重指数	舒张压	口服耐糖量测试	胰岛素释放实验	肱三头肌皮褶厚度
count	1000.000000	1000.000000	1000.000000	1000.000000	951.000000	1000.000000	1000.000000	1000.000000
mean	500.500000	0.481000	1986.386000	39.439000	89.638275	5.872314	4.102700	7.064240
std	288.819436	0.499889	8.816163	11.284861	9.379124	1.930880	8.594005	13.900938
min	1.000000	0.000000	1958.000000	0.000000	28.000000	-1.000000	0.000000	0.000000
25%	250.750000	0.000000	1979.000000	29.975000	85.000000	4.516000	0.000000	0.000000
50%	500.500000	0.000000	1987.000000	38.900000	89.000000	5.851500	0.000000	0.000000
75%	750.250000	1.000000	1994.000000	48.950000	96.000000	7.465000	7.202500	3.820000
max	1000.000000	1.000000	2003.000000	60.000000	112.000000	10.613000	123.890000	44.900000

2. 数据探索及预处理

2.1 缺失值

统计每个字段的缺失比例，并进行填充。可以看到舒张压指标的缺失值较多，用字段均值将其填充。

print('训练集各字段缺失比例：')
print(train_data.isnull().mean(0))
print('\n测试集各字段缺失比例：')
print(test_data.isnull().mean(0))

# 用均值填充缺失值
train_data['舒张压'] = train_data['舒张压'].fillna(train_data['舒张压'].mean())
test_data['舒张压'] = test_data['舒张压'].fillna(test_data['舒张压'].mean())

训练集各字段缺失比例：
编号          0.000000
性别          0.000000
出生年份        0.000000
体重指数        0.000000
糖尿病家族史      0.000000
舒张压         0.048718
口服耐糖量测试     0.000000
胰岛素释放实验     0.000000
肱三头肌皮褶厚度    0.000000
患有糖尿病标识     0.000000
dtype: float64

测试集各字段缺失比例：
编号          0.000
性别          0.000
出生年份        0.000
体重指数        0.000
糖尿病家族史      0.000
舒张压         0.049
口服耐糖量测试     0.000
胰岛素释放实验     0.000
肱三头肌皮褶厚度    0.000
dtype: float64

2.2 分析字段类型

print(train_data.columns)

train_data.describe()

Index(['编号', '性别', '出生年份', '体重指数', '糖尿病家族史', '舒张压', '口服耐糖量测试', '胰岛素释放实验',
       '肱三头肌皮褶厚度', '患有糖尿病标识'],
      dtype='object')

	编号	性别	出生年份	体重指数	舒张压	口服耐糖量测试	胰岛素释放实验	肱三头肌皮褶厚度	患有糖尿病标识
count	5070.000000	5070.000000	5070.000000	5070.000000	5070.000000	5070.000000	5070.000000	5070.000000	5070.000000
mean	2535.500000	0.456805	1986.869231	37.986785	89.423595	5.612839	4.114321	6.994371	0.381854
std	1463.727263	0.498180	8.919737	11.447095	9.038394	2.257649	8.726001	13.651442	0.485889
min	1.000000	0.000000	1943.000000	0.000000	30.000000	-1.000000	0.000000	0.000000	0.000000
25%	1268.250000	0.000000	1980.000000	28.400000	85.000000	4.314000	0.000000	0.000000	0.000000
50%	2535.500000	0.000000	1987.000000	36.550000	89.000000	5.760000	0.000000	0.000000	0.000000
75%	3802.750000	1.000000	1995.000000	47.600000	95.000000	7.193000	7.100000	4.120000	1.000000
max	5070.000000	1.000000	2009.000000	65.900000	126.000000	10.839000	108.960000	45.000000	1.000000

编号与是否患病没关系，删除;

性别为类别变量，只有0，1，不再需要进行编码；

糖尿病家族病史为文本型变量，需要转化为数值变量；

其他均为数值型变量，可以暂时不变。

train_data = train_data.drop(['编号'], axis=1)
test_data = test_data.drop(['编号'], axis=1)

2.3 计算字段相关性

查看各字段之间的相关性，防止多重共线性。

Ref:
Python绘制相关性热力图

train_corr = train_data.drop('糖尿病家族史',axis=1).corr()

import seaborn as sns

plt.subplots(figsize=(9,9),dpi=80,facecolor='w') # 设置画布大小，分辨率，和底色
plt.rcParams['font.sans-serif'] = ['SimHei'] # 黑体
plt.rcParams['axes.unicode_minus'] = False   # 解决无法显示符号的问题
sns.set(font='SimHei', font_scale=0.8)       # 解决Seaborn中文显示问题

#annot为热力图上显示数据；fmt='.2g'为数据保留两位有效数字,square呈现正方形，vmax最大值为1
fig=sns.heatmap(train_corr,annot=True, vmax=1, square=True, cmap="Blues", fmt='.2g')

#保存图片
fig.get_figure().savefig('train_corr.png',bbox_inches='tight',transparent=True)
#bbox_inches让图片显示完整，transparent=True让图片背景透明

这里中文显示出了点问题，但是可以看到各特征变量与是否患病没有显著线性关系（但可能有非线性关系），各特征变量之间也不存在多重共线性，可以继续下一步操作。

3. 特征工程

这一步至关重要，主要是有两个目的：

• 特征构造：尝试构建有价值的新变量；
• 特征筛选：删除对因变量影响不大的冗余变量。

由于这里的特征也不是很多，就不做筛选了。

3.1 特征构造

可以用统计指标，已有知识、经验构造新的变量，具体到这个问题上可以有BMI指数、舒张压范围、年龄等。

特征构造方法：

1. 特征的统计指标；
2. 特征之间的四则运算；
3. 交叉特征；
4. 分解类别特征。如将三个颜色分解为“知道颜色”和“不知道颜色”。
5. 特征分箱。将数值型特征变量按段划分，得到类别型特征。
6. 重构特征。单位转换、整数部分与小数部分分离等。
7. 根据已有经验构造新的特征变量，比如xx因子。

Ref:
[1] 深度了解特征工程

# 将出生年份换算成年龄
train_data['年龄']=2022-train_data['出生年份']  #换成年龄
test_data['年龄']=2022-test_data['出生年份']

train_data = train_data.drop('出生年份', axis=1)
test_data = test_data.drop('出生年份', axis=1)

# 家族史转换， 方法一，label编码
from sklearn.preprocessing import OneHotEncoder, LabelEncoder

def FHOD(a):
   if a=='无记录':
       return 0
   elif a=='叔叔或者姑姑有一方患有糖尿病' or a=='叔叔或姑姑有一方患有糖尿病':
       return 1
   else:
       return 2

train_data['糖尿病家族史'] = train_data['糖尿病家族史'].apply(FHOD)
test_data['糖尿病家族史'] = test_data['糖尿病家族史'].apply(FHOD)

# history = train_data['糖尿病家族史']
# print(set(history))

# history.loc[history=='叔叔或姑姑有一方患有糖尿病']  = '叔叔或者姑姑有一方患有糖尿病'

# le = LabelEncoder()
# h = le.fit_transform(history)

# 方法二，onehot 编码
# def onehot_transform(data):
#     # 将家族史的文本型变量转换为onehot编码。

#     onehot = OneHotEncoder()
#     data.loc[data['糖尿病家族史']=='叔叔或姑姑有一方患有糖尿病', '糖尿病家族史']  = '叔叔或者姑姑有一方患有糖尿病'

#     data_onehot = pd.DataFrame(onehot.fit_transform(data[['糖尿病家族史']]).toarray(),
#     columns=onehot.get_feature_names(['糖尿病家族史']), dtype='int32')

#     return data_onehot

# data_train_history = onehot_transform(train_data)
# data_test_history = onehot_transform(test_data)

def BMI(a):
    """
    人体的成人体重指数正常值是在18.5-24之间
    低于18.5是体重指数过轻
    在24-27之间是体重超重
    27以上考虑是肥胖
    高于32了就是非常的肥胖。
    """
    if a<18.5:
        return 0
    elif 18.5<=a<=24:
        return 1
    elif 24<a<=27:
        return 2
    elif 27<a<=32:
        return 3
    else:
        return 4

train_data['BMI']=train_data['体重指数'].apply(BMI)
test_data['BMI']=test_data['体重指数'].apply(BMI)

# 转换舒张压为类别型变量
def DBP(a):
    # 舒张压范围为60-90
    if a<60:
        return 0
    elif 60<=a<=90:
        return 1
    elif a>90:
        return 2
    else:
        return a

train_data['DBP'] = train_data['舒张压'].apply(DBP)
test_data['DBP'] = test_data['舒张压'].apply(DBP)

X_train = train_data.drop('患有糖尿病标识', axis=1)
Y_train = train_data['患有糖尿病标识']
X_train['年龄'] = X_train['年龄'].astype(float)
X_test = test_data

train_data.describe()

	性别	体重指数	糖尿病家族史	舒张压	口服耐糖量测试	胰岛素释放实验	肱三头肌皮褶厚度	患有糖尿病标识	年龄	BMI	DBP
count	5070.000000	5070.000000	5070.000000	5070.000000	5070.000000	5070.000000	5070.000000	5070.000000	5070.000000	5070.000000	5070.000000
mean	0.456805	37.986785	0.601183	89.423595	5.612839	4.114321	6.994371	0.381854	35.130769	3.301972	1.394477
std	0.498180	11.447095	0.764882	9.038394	2.257649	8.726001	13.651442	0.485889	8.919737	1.051700	0.510116
min	0.000000	0.000000	0.000000	30.000000	-1.000000	0.000000	0.000000	0.000000	13.000000	0.000000	0.000000
25%	0.000000	28.400000	0.000000	85.000000	4.314000	0.000000	0.000000	0.000000	27.000000	3.000000	1.000000
50%	0.000000	36.550000	0.000000	89.000000	5.760000	0.000000	0.000000	0.000000	35.000000	4.000000	1.000000
75%	1.000000	47.600000	1.000000	95.000000	7.193000	7.100000	4.120000	1.000000	42.000000	4.000000	2.000000
max	1.000000	65.900000	2.000000	126.000000	10.839000	108.960000	45.000000	1.000000	79.000000	4.000000	2.000000

4. 模型训练

可用于分类问题的模型非常丰富，常见的如下图：

4.1 LightGBM (0.96206)

首先尝试构建LightGBM模型。

Ref:
[1] Lightgbm原理、参数详解及python实例
[2] 深入理解LightGBM

以下为lightgbm采用5折交叉训练的代码：

#使用Lightgbm方法训练数据集，使用5折交叉验证的方法获得5个测试集预测结果
from sklearn.model_selection import KFold
from sklearn.model_selection import StratifiedKFold, GridSearchCV

def select_by_lgb(train_data,train_label,test_data,random_state=1234, n_splits=5,metric='auc',num_round=10000,early_stopping_rounds=100):
#    kfold = KFold(n_splits=n_splits, shuffle=True, random_state=random_state)
    kfold = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=random_state)
    fold=0
    result0 = []
    for train_idx, val_idx in kfold.split(train_data, train_label):
        random_state+=1
        train_x = train_data.loc[train_idx]
        train_y = train_label.loc[train_idx]
        test_x = train_data.loc[val_idx]
        test_y = train_label.loc[val_idx]
        clf = lightgbm
        train_matrix=clf.Dataset(train_x,label=train_y)
        test_matrix=clf.Dataset(test_x,label=test_y)
        params={
                'boosting_type': 'gbdt',
                'objective': 'binary',
                'learning_rate': 0.1,
                # 'max_depth': 7,
                # 'num_leaves': 10,
                'metric': metric,
                'seed': random_state,
                'silent': True,
                'nthread':-1 }
        model=clf.train(params,train_matrix,num_round,valid_sets=test_matrix,early_stopping_rounds=early_stopping_rounds)
        pre_y=model.predict(test_data)
        result0.append(pre_y)
        fold+=1

        pred_test = pd.DataFrame(result0).T

        # 将5次预测结果求平均值
        pred_test['average'] = pred_test.mean(axis=1)

        #因为竞赛需要你提交最后的预测判断，而模型给出的预测结果是概率，因此我们认为概率>0.5的即该患者有糖尿病，概率<=0.5的没有糖尿病
        pred_test['label'] = pred_test['average'].apply(lambda x:1 if x>0.5 else 0)

        ## 导出结果
        result = pd.read_csv('提交示例.csv')
        result['label']=pred_test['label']

    return result

后面其他模型也需要进行k折交叉训练，这里定义一个k折交叉训练的函数，方便后续调用。

from sklearn.model_selection import KFold, StratifiedKFold
from sklearn.metrics import roc_auc_score, f1_score

def SKFold(train_data,train_label,test_data, model, random_state=1234, n_splits=5,metric='auc',num_round=10000,early_stopping_rounds=100):
    # 采用分层K折交叉验证训练模型。
    kfold = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=random_state)
    fold = 1
    pred_test = []
    for train_idx, val_idx in kfold.split(train_data, train_label):
        random_state+=1
        train_x = train_data.loc[train_idx]
        train_y = train_label.loc[train_idx]
        val_x = train_data.loc[val_idx]
        val_y = train_label.loc[val_idx]
        eval_set = (val_x, val_y)
        clf = model
        model_trained = clf.fit(train_x, train_y)
        # model_trained = clf.fit(train_x,train_y,early_stopping_rounds=early_stopping_rounds, verbose=False)
        # model_trained = clf.fit(train_x, train_y, eval_set=eval_set, early_stopping_rounds=early_stopping_rounds)
        pre_y = model_trained.predict(test_data)
        pred_test.append(pre_y)

        auc_train = roc_auc_score(train_y, model_trained.predict(train_x))
        auc_val = roc_auc_score(val_y, model_trained.predict(val_x))
        f_score_train = f1_score(train_y, model_trained.predict(train_x))
        f_score_val = f1_score(val_y, model_trained.predict(val_x))
        print('Fold: %d, AUC_train: %.4f, AUC_val: %.4f, F1-score_train: %.4f, F1-score_val: %.4f'%(fold, 
            auc_train, auc_val, f_score_train, f_score_val))
        fold += 1

    pred_test = pd.DataFrame(pred_test).T
    # 将5次预测结果求平均值
    pred_test['average'] = pred_test.mean(axis=1)

    #因为竞赛需要你提交最后的预测判断，而模型给出的预测结果是概率，因此我们认为概率>0.5的即该患者有糖尿病，概率<=0.5的没有糖尿病
    pred_test['label'] = pred_test['average'].apply(lambda x:1 if x>0.5 else 0)

    ## 导出结果
    result=pd.read_csv('提交示例.csv')
    result['label']=pred_test['label']

    return result

由于比赛的测试集未公布，我们只能提交预测结果然后得到测试集上的分数，这里以表现较好的lightgbm模型作为baseline，若与lightgbm的预测结果相差较多则说明该模型表现不行。

def evaluate(result_LightGBM, result_others):
    # 以lightGBM的结果为基准，评估其他模型的表现。

    c = result_LightGBM['label'] - result_others['label']

    count = 0
    for i in c:
        if i != 0:
            count += 1

    print('与LightGBM预测不同的样本数： ', count)
    print(c[c!=0])

    return count

先用select_by_lgb快速跑出一个baseline，然后用网格搜索得到最优参数，接着用最优参数组合在训练一遍模型，最后将结果提交。

random_state = 1234

result_LightGBM = select_by_lgb(X_train, Y_train, X_test)   #baseline
result_LightGBM.to_csv('result_lightGBM.csv',index=False)

# 试试网格搜索最优参数
import lightgbm as lgb
params_test = {
    'max_depth': range(4, 10, 1),
    'num_leaves': range(10, 60, 10)
}

skf = StratifiedKFold(n_splits=5)  

gsearch1 = GridSearchCV(estimator=lgb.LGBMClassifier(boosting_type='gbdt',objective='binary',metrics='auc',learning_rate=0.1, n_estimators=325, 
    max_depth=8, bagging_fraction = 0.8,feature_fraction = 0.8),  param_grid=params_test,
        scoring='roc_auc', cv=skf, n_jobs=-1)

gsearch1.fit(X_train, Y_train)
print(gsearch1.best_params_)
print(gsearch1.best_score_)

# 用最优参数再训练一遍
model_lgb = lgb.LGBMClassifier(boosting_type='gbdt',objective='binary',metrics='auc',
    learning_rate=0.1, n_estimators=200, num_leaves=10, silent=True,
    max_depth=7)

result_SKFold_lgb = SKFold(X_train, Y_train, X_test, model_lgb, n_splits=5)
result_SKFold_lgb.to_csv('result_SKFold_lgb.csv',index=False)

diff_lgb = evaluate(result_LightGBM, result_SKFold_lgb)

[LightGBM] [Warning] Unknown parameter: silent
[LightGBM] [Warning] Unknown parameter: silent
[LightGBM] [Info] Number of positive: 1549, number of negative: 2507
[LightGBM] [Warning] Auto-choosing row-wise multi-threading, the overhead of testing was 0.000386 seconds.
You can set `force_row_wise=true` to remove the overhead.
And if memory is not enough, you can set `force_col_wise=true`.
[LightGBM] [Info] Total Bins 1047
[LightGBM] [Info] Number of data points in the train set: 4056, number of used features: 10
[LightGBM] [Warning] Unknown parameter: silent
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.381903 -> initscore=-0.481477
[LightGBM] [Info] Start training from score -0.481477
[1]	valid_0's auc: 0.979322
Training until validation scores don't improve for 100 rounds
[2]	valid_0's auc: 0.980354
[3]	valid_0's auc: 0.982351
[4]	valid_0's auc: 0.981993

Early stopping, best iteration is:
[40]	valid_0's auc: 0.989851
[LightGBM] [Warning] feature_fraction is set=0.8, colsample_bytree=1.0 will be ignored. Current value: feature_fraction=0.8
[LightGBM] [Warning] bagging_fraction is set=0.8, subsample=1.0 will be ignored. Current value: bagging_fraction=0.8
{'max_depth': 7, 'num_leaves': 10}
0.9902553653233458
Fold: 1, AUC_train: 0.9917, AUC_val: 0.9461, F1-score_train: 0.9909, F1-score_val: 0.9358
[LightGBM] [Warning] feature_fraction is set=0.8, colsample_bytree=1.0 will be ignored. Current value: feature_fraction=0.8
[LightGBM] [Warning] bagging_fraction is set=0.8, subsample=1.0 will be ignored. Current value: bagging_fraction=0.8
Fold: 2, AUC_train: 0.9852, AUC_val: 0.9487, F1-score_train: 0.9837, F1-score_val: 0.9386
[LightGBM] [Warning] feature_fraction is set=0.8, colsample_bytree=1.0 will be ignored. Current value: feature_fraction=0.8
[LightGBM] [Warning] bagging_fraction is set=0.8, subsample=1.0 will be ignored. Current value: bagging_fraction=0.8
Fold: 3, AUC_train: 0.9863, AUC_val: 0.9457, F1-score_train: 0.9853, F1-score_val: 0.9328
[LightGBM] [Warning] feature_fraction is set=0.8, colsample_bytree=1.0 will be ignored. Current value: feature_fraction=0.8
[LightGBM] [Warning] bagging_fraction is set=0.8, subsample=1.0 will be ignored. Current value: bagging_fraction=0.8
Fold: 4, AUC_train: 0.9876, AUC_val: 0.9607, F1-score_train: 0.9860, F1-score_val: 0.9490
[LightGBM] [Warning] feature_fraction is set=0.8, colsample_bytree=1.0 will be ignored. Current value: feature_fraction=0.8
[LightGBM] [Warning] bagging_fraction is set=0.8, subsample=1.0 will be ignored. Current value: bagging_fraction=0.8
Fold: 5, AUC_train: 0.9858, AUC_val: 0.9505, F1-score_train: 0.9841, F1-score_val: 0.9403
与LightGBM预测不同的样本数：  7
24     1
35    -1
43     1
76     1
434    1
442    1
501   -1
Name: label, dtype: int64

4.2 随机森林（0.96324）

Ref:
[1] Permutation Importance vs Random Forest Feature Importance (MDI)

from sklearn.ensemble import RandomForestClassifier

forest = RandomForestClassifier(max_depth=5, random_state=1234)
forest.fit(X_train, Y_train)
pred_forest = forest.predict(X_test)

result=pd.read_csv('提交示例.csv')
result['label']=pred_forest
result.to_csv('result_RandomForest.csv',index=False)

feature_importance_forest = pd.Series(forest.feature_importances_, 
            index=X_train.columns).sort_values(ascending=True)

plt.figure(figsize=(10, 7), dpi=80)

ax = feature_importance_forest.plot.barh()
ax.set_title("Random Forest Feature Importances (MDI)")
# ax.figure.tight_layout()

## 网格搜索最优参数组合

params_test = {
    'max_depth': range(3, 20, 2),
    'n_estimators': range(100, 600, 100),
    'min_samples_leaf': [2, 4, 6]
}

skf = StratifiedKFold(n_splits=5)  

gsearch2 = GridSearchCV(estimator=RandomForestClassifier(n_estimators=200, max_depth=5, random_state=random_state), 
     param_grid=params_test, scoring='roc_auc', cv=skf, n_jobs=-1)

gsearch2.fit(X_train, Y_train)
print(gsearch2.best_params_)
print(gsearch2.best_score_)

{'max_depth': 13, 'min_samples_leaf': 2, 'n_estimators': 400}
0.9927105570462718

model_forest = RandomForestClassifier(n_estimators=400, max_depth=13, random_state=random_state)
result_SKFold_forest = SKFold(X_train, Y_train, X_test, model_forest) 
result_SKFold_forest.to_csv('result_skfold_RandomForest.csv',index=False)

diff_skold_forest = evaluate(result_LightGBM, result_SKFold_forest)

Fold: 1, AUC_train: 0.9935, AUC_val: 0.9516, F1-score_train: 0.9935, F1-score_val: 0.9424
Fold: 2, AUC_train: 0.9910, AUC_val: 0.9564, F1-score_train: 0.9909, F1-score_val: 0.9468
Fold: 3, AUC_train: 0.9919, AUC_val: 0.9490, F1-score_train: 0.9919, F1-score_val: 0.9396
Fold: 4, AUC_train: 0.9919, AUC_val: 0.9647, F1-score_train: 0.9919, F1-score_val: 0.9551
Fold: 5, AUC_train: 0.9913, AUC_val: 0.9591, F1-score_train: 0.9912, F1-score_val: 0.9497
与LightGBM预测不同的样本数：  11
8      1
35    -1
47    -1
52    -1
60     1
64     1
76     1
78     1
85     1
618   -1
796   -1
Name: label, dtype: int64

4.3 XGBoost (0.95981)

Ref:
[1] XGBoost：在Python中使用XGBoost
[2] Python机器学习笔记：XgBoost算法
[3] python包xgboost安装和简单使用
[4] 深入理解XGBoost，优缺点分析，原理推导及工程实现
[5] XGBoost的原理、公式推导、Python实现和应用
[6] XGBoost官方文档

import xgboost as xgb
from xgboost import plot_importance
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import StratifiedKFold

import warnings
warnings.filterwarnings('ignore')

# 分层k折交叉检验
skf = StratifiedKFold(n_splits=5)  

result_xgb = []
fold = 1
for train_idx, val_idx in skf.split(X_train, Y_train):
    train_x = X_train.loc[train_idx]
    train_y = Y_train.loc[train_idx]
    val_x = X_train.loc[val_idx]
    val_y = Y_train.loc[val_idx]
    d_train = xgb.DMatrix(train_x, train_y)
    d_val = xgb.DMatrix(val_x, val_y)
    d_test = xgb.DMatrix(X_test)

    params = {
        'max_depth':5,
        'min_child_weight':1,
        'num_class':2,
        'eta': 0.1,  #学习率
        'gamma': 0.1, #后剪枝参数，取值在[0, 1]，越大越保守
        'seed': 1234,
        'alpha': 1,  #L1正则项的惩罚系数
        'eval_metric': 'auc'
    }
    num_round = 500

    # # 方式一：采用sklearn接口，采用fit 和 predict
    # model_xgb = xgb.XGBClassifier()
    # model_xgb.fit(train_x, train_y, verbose=False) 
    # pred_train = model_xgb.predict(train_x)
    # pred_val = model_xgb.predict(val_x)
    # pred_xgb = model_xgb.predict(X_test)

    # 方式二：采用xgboost原生接口，采用train和predict，方便调参
    model_xgb = xgb.train(params, d_train, num_round)
    pred_train = model_xgb.predict(d_train)
    pred_val = model_xgb.predict(d_val)
    pred_xgb = model_xgb.predict(d_test)

    auc_train = roc_auc_score(train_y, pred_train)
    auc_val = roc_auc_score(val_y, pred_val)
    f_score_train = f1_score(train_y, pred_train)
    f_score_val = f1_score(val_y, pred_val)
    print('Fold: %d, AUC_train: %.4f, AUC_val: %.4f, F1-score_train: %.4f, F1-score_val: %.4f'%(fold, 
        auc_train, auc_val, f_score_train, f_score_val))

    result_xgb.append(pred_xgb)

    fold += 1

result_xgb = pd.DataFrame(result_xgb).T
print('result_xgb.shape = ', result_xgb.shape)

# 将5次预测结果求平均值
result_xgb['average'] = result_xgb.mean(axis=1)

# 最终预测结果
result_xgb['label'] = result_xgb['average'].apply(lambda x:1 if x>0.5 else 0)

# 特征重要性
plot_importance(model_xgb)
plt.show()

# 导出结果
result = pd.read_csv('提交示例.csv')
result['label'] = result_xgb['label']
result.to_csv('result_XGBoost_StratifiedKFold.csv',index=False)

diff_xgb = evaluate(result_LightGBM, result_xgb)

Fold: 1, AUC_train: 0.9935, AUC_val: 0.9463, F1-score_train: 0.9929, F1-score_val: 0.9349
Fold: 2, AUC_train: 0.9952, AUC_val: 0.9556, F1-score_train: 0.9948, F1-score_val: 0.9456
Fold: 3, AUC_train: 0.9952, AUC_val: 0.9518, F1-score_train: 0.9948, F1-score_val: 0.9415
Fold: 4, AUC_train: 0.9906, AUC_val: 0.9524, F1-score_train: 0.9893, F1-score_val: 0.9407
Fold: 5, AUC_train: 0.9924, AUC_val: 0.9573, F1-score_train: 0.9919, F1-score_val: 0.9482
result_xgb.shape =  (1000, 5)

与LightGBM预测不同的样本数：  6
21   -1
24    1
28   -1
35   -1
43    1
74   -1
Name: label, dtype: int64

4.4 CatBoost（0.95854）

CatbBoost 是GBDT算法框架的一种改进实现，其主要创新点有：

• 支持类别性变量。嵌入了自动将类别型特征处理为数值型特征的创新算法。
• 使用了组合类别特征，丰富特征维度。
• 采用排序提升的方法对抗训练集中的噪声点，从而避免梯度估计的偏差，进而解决预测偏移的问题。
• 采用了完全对称树作为基模型。

Ref:
[1] 深入理解CatBoost
[2] Catboost 一个超级简单实用的boost算法

import catboost as cb
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import StratifiedKFold

# 分层k折交叉检验
skf = StratifiedKFold(n_splits=5)  

categorical_features_index = np.where(X_train.dtypes != float)[0]
print(X_train.columns[categorical_features_index])

result_cat = []
fold = 1
for train_idx, val_idx in skf.split(X_train, Y_train):
    train_x = X_train.loc[train_idx]
    train_y = Y_train.loc[train_idx]
    val_x = X_train.loc[val_idx]
    val_y = Y_train.loc[val_idx]

    model_catboost = cb.CatBoostClassifier(eval_metric='AUC', cat_features=categorical_features_index, 
        depth=6, n_estimators=400, learning_rate=0.5, verbose=False)
    model_catboost.fit(train_x, train_y, eval_set=(val_x, val_y), plot=False)

    pred_train = model_catboost.predict(train_x)
    pred_val = model_catboost.predict(val_x)

    auc_train = roc_auc_score(train_y, pred_train)
    auc_val = roc_auc_score(val_y, pred_val)
    f_score_train = f1_score(train_y, pred_train)
    f_score_val = f1_score(val_y, pred_val)
    print('Fold: %d, AUC_train: %.4f, AUC_val: %.4f, F1-score_train: %.4f, F1-score_val: %.4f'%(fold, 
        auc_train, auc_val, f_score_train, f_score_val))

    pred_catboost = model_catboost.predict(X_test)
    result_cat.append(pred_catboost)

    fold += 1

result_cat = pd.DataFrame(result_cat).T
print('result_cat.shape = ', result_cat.shape)

# 将5次预测结果求平均值
result_cat['average'] = result_cat.mean(axis=1)

# 最终预测结果
result_cat['label'] = result_cat['average'].apply(lambda x:1 if x>0.5 else 0)

# 导出结果
result = pd.read_csv('提交示例.csv')
result['label'] = result_cat['label']
result.to_csv('result_CatBoost_StratifiedKFold.csv',index=False)

diff_catboost = evaluate(result_LightGBM, result_cat)

feature_importance_catboost = model_catboost.feature_importances_

plt.figure(figsize=(10,8), dpi=80)
plt.barh(col_names, feature_importance_catboost)
plt.show()

Index(['性别', '糖尿病家族史', 'BMI', 'DBP'], dtype='object')
Fold: 1, AUC_train: 0.9902, AUC_val: 0.9419, F1-score_train: 0.9887, F1-score_val: 0.9305
Fold: 2, AUC_train: 0.9775, AUC_val: 0.9608, F1-score_train: 0.9722, F1-score_val: 0.9510
Fold: 3, AUC_train: 0.9988, AUC_val: 0.9543, F1-score_train: 0.9984, F1-score_val: 0.9442
Fold: 4, AUC_train: 0.9590, AUC_val: 0.9448, F1-score_train: 0.9498, F1-score_val: 0.9298
Fold: 5, AUC_train: 0.9651, AUC_val: 0.9484, F1-score_train: 0.9578, F1-score_val: 0.9377
result_cat.shape =  (1000, 5)
与LightGBM预测不同的样本数：  16
8      1
33    -1
35    -1
47    -1
52    -1
60     1
64     1
74    -1
76     1
83    -1
85     1
89    -1
166   -1
501   -1
618   -1
796   -1
Name: label, dtype: int64

4.5 AdaBoost（0.96098）

from sklearn.ensemble import AdaBoostClassifier
from sklearn.tree import DecisionTreeClassifier

random_state = 1234
model_tree = DecisionTreeClassifier(max_depth=5, random_state=random_state)
model_adaboost = AdaBoostClassifier(base_estimator=model_tree, n_estimators=200, 
                    random_state=random_state)

result_adaboost = SKFold(X_train, Y_train, X_test, model_adaboost) 
result_adaboost.to_csv('result_AdaBoost.csv',index=False)

# 评估
diff_skold_adaboost = evaluate(result_LightGBM, result_adaboost)

# 特征重要性
feature_importance_adaboost = model_adaboost.feature_importances_

plt.figure(figsize=(10,8), dpi=80)
plt.rc('font', size = 18)
plt.barh(col_names, feature_importance_adaboost)
plt.title('Feature importances computed by AdaBoost')
plt.show()

Fold: 1, AUC_train: 1.0000, AUC_val: 0.9458, F1-score_train: 1.0000, F1-score_val: 0.9347
Fold: 2, AUC_train: 1.0000, AUC_val: 0.9445, F1-score_train: 1.0000, F1-score_val: 0.9333
Fold: 3, AUC_train: 1.0000, AUC_val: 0.9380, F1-score_train: 1.0000, F1-score_val: 0.9263
Fold: 4, AUC_train: 1.0000, AUC_val: 0.9488, F1-score_train: 1.0000, F1-score_val: 0.9357
Fold: 5, AUC_train: 1.0000, AUC_val: 0.9536, F1-score_train: 1.0000, F1-score_val: 0.9432
与LightGBM预测不同的样本数：  12
35    -1
47    -1
50    -1
52    -1
60     1
64     1
76     1
85     1
92    -1
94     1
495    1
796   -1
Name: label, dtype: int64

4.6 集成模型（0.95971）

挑几个表现较好的模型进行集成。

%%time
skf = StratifiedKFold(n_splits=5)  

categorical_features_index = np.where(X_train.dtypes != float)[0]
print('类别型特征： ', X_train.columns[categorical_features_index])
cat_features = list(map(lambda x:int(x), categorical_features_index))
random_state = 1234

fold = 1
for train_idx, val_idx in skf.split(X_train, Y_train):
    train_x = X_train.loc[train_idx]
    train_y = Y_train.loc[train_idx]
    val_x = X_train.loc[val_idx]
    val_y = Y_train.loc[val_idx]
    d_train = xgb.DMatrix(train_x, train_y)
    d_val = xgb.DMatrix(val_x, val_y)
    d_test = xgb.DMatrix(X_test)

    params_xgb = {
        'max_depth':5,
        'min_child_weight':1,
        'num_class':2,
        'eta': 0.1,  #学习率
        'gamma': 0.1, #后剪枝参数，取值在[0, 1]，越大越保守
        'seed': 1234,
        'alpha': 1,  #L1正则项的惩罚系数
        'eval_metric': 'auc'
    }

    num_round = 500
    early_stopping_rounds = 100

    model_lightGBM = lgb.LGBMClassifier(boosting_type='gbdt',objective='binary',metrics='auc',
        learning_rate=0.1, n_estimators=200, num_leaves=10, silent=True,
        max_depth=7)
    model_lightGBM.fit(X_train, Y_train)

    model_forest = RandomForestClassifier(max_depth=13, n_estimators=400, random_state=1234)
    model_forest.fit(X_train, Y_train)

    model_tree = DecisionTreeClassifier(max_depth=5, random_state=random_state)
    model_adaboost = AdaBoostClassifier(base_estimator=model_tree, n_estimators=200, 
                        random_state=random_state)
    model_adaboost.fit(X_train, Y_train)

    model_xgb = xgb.train(params_xgb, d_train, num_round)

    model_catboost = cb.CatBoostClassifier(eval_metric='AUC', cat_features=categorical_features_index, 
        depth=6, iterations=400, learning_rate=0.5, verbose=False)
    model_catboost.fit(train_x, train_y, eval_set=(val_x, val_y), plot=False)

    print('Fold: %d finished training. '%fold)
    fold += 1

pred_lightGBM = model_lightGBM.predict(test_data)
# pred_lightGBM = list(map(lambda x: 1 if x>0.5 else 0, pred_lightGBM))  #调用lightGBM原生接口时使用
pred_forest = forest.predict(X_test)
pred_adaboost = model_adaboost.predict(X_test)
pred_xgb = model_xgb.predict(d_test)
pred_catboost = model_catboost.predict(X_test)
pred_all = pd.DataFrame({'lightGBM': pred_lightGBM,
                            'RandomForest': pred_forest,
                            'AdaBoost': pred_adaboost,
                            'XGBoost': pred_xgb,
                            'CatBoost': pred_catboost})

pred_all['Average'] = pred_all.mean(axis=1)

# 最终预测结果
pred_all['label'] = pred_all['Average'].apply(lambda x:1 if x>0.5 else 0)

# 导出结果
result = pd.read_csv('提交示例.csv')
result['label'] = pred_all['label']
result.to_csv('result_Ensemble.csv',index=False)

diff_ensemble = evaluate(result_LightGBM, result)

类别型特征：  Index(['性别', '糖尿病家族史', 'BMI', 'DBP'], dtype='object')


Custom logger is already specified. Specify more than one logger at same time is not thread safe.

Fold: 1 finished training. 
Fold: 2 finished training. 
Fold: 3 finished training. 
Fold: 4 finished training. 
Fold: 5 finished training. 
与LightGBM预测不同的样本数：  11
24     1
33    -1
35    -1
43     1
52    -1
60     1
64     1
76     1
78     1
85     1
796   -1
Name: label, dtype: int64
Wall time: 1min 31s

4.7 Stacking（0.96577）

Stacking的思想为在初始数据集上训练若干个基学习器，并将这几个基学习器的预测结果作为新的训练集，来训练一个新的学习器，并将其预测结果作为最终输出。

Stacking本质是一种层级结构，第一层有n个基学习器，每个基学习器进行k折交叉训练，把每一折的验证集（validation set）的预测结果输出并拼接在一起，把这n个模型的训练集预测结果作为新的训练集，将这n个模型的测试集预测结果拼接在一起作为新的测试集。

（图源见水印）

Ref:
[1] stacking模型融合
[2] Kaggle上分技巧——单模K折交叉验证训练+多模型融合

model_tree = DecisionTreeClassifier(max_depth=5, random_state=random_state)

clfs = [lgb.LGBMClassifier(boosting_type='gbdt',objective='binary',metrics='auc',
        learning_rate=0.1, n_estimators=200, num_leaves=10, silent=True,
        max_depth=7),
        RandomForestClassifier(max_depth=13, n_estimators=400, random_state=1234),
        AdaBoostClassifier(base_estimator=model_tree, n_estimators=200, 
                        random_state=random_state),
        xgb.XGBClassifier(),
        cb.CatBoostClassifier(eval_metric='AUC', cat_features=categorical_features_index, 
        depth=6, iterations=400, learning_rate=0.5, verbose=False)]

data_train = np.zeros((X_train.shape[0], len(clfs)))
data_test = np.zeros((X_test.shape[0], len(clfs)))

# 5折stacking
n_splits = 5
skf = StratifiedKFold(n_splits)

# 第一层，训练各个个体学习器
for i, clf in enumerate(clfs):
    # 依次训练各个模型
    d_test = np.zeros((X_test.shape[0], n_splits)) #存放个体学习器在测试集上的预测输出
    for fold, (train_idx, val_idx) in enumerate(skf.split(X_train, Y_train)):
        #5折交叉训练，第j折拿来预测并作为第二层模型的训练集，剩余部分拿来训练模型。
        train_x = X_train.loc[train_idx]
        train_y = Y_train.loc[train_idx]
        val_x = X_train.loc[val_idx]
        val_y = Y_train.loc[val_idx]
        
        clf.fit(train_x, train_y)
        pred_train = clf.predict(train_x)
        pred_val = clf.predict(val_x)
        data_train[val_idx, i] = pred_val
        d_test[:, fold] = clf.predict(X_test)

        auc_train = roc_auc_score(train_y, pred_train)
        auc_val = roc_auc_score(val_y, pred_val)
        f_score_train = f1_score(train_y, pred_train)
        f_score_val = f1_score(val_y, pred_val)
        print('Classifier:%d, Fold: %d, AUC_train: %.4f, AUC_val: %.4f, F1-score_train: %.4f, F1-score_val: %.4f'%(i+1,
            fold+1, auc_train, auc_val, f_score_train, f_score_val))

    #对于测试集，直接用这交叉验证训练的每个模型的预测值均值作为新的特征
    data_test[:, i] = d_test.mean(axis=1)


data_train = pd.DataFrame(data_train)
data_test = pd.DataFrame(data_test)

# 第二层改用高级点的模型，并进行5折交叉训练
# model_forest = RandomForestClassifier(max_depth=5, random_state=1234)
model_2 = lgb.LGBMClassifier(boosting_type='gbdt',objective='binary',metrics='auc',
        learning_rate=0.3, n_estimators=200, num_leaves=10, silent=True,
        max_depth=7)
result_stack = SKFold(data_train, Y_train, data_test, model_2)
result_stack.to_csv('result_stack.csv', index=False)

diff_stack = evaluate(result_LightGBM, result_stack)

Classifier:1, Fold: 1, AUC_train: 0.9894, AUC_val: 0.9465, F1-score_train: 0.9880, F1-score_val: 0.9340
Classifier:1, Fold: 2, AUC_train: 0.9883, AUC_val: 0.9554, F1-score_train: 0.9870, F1-score_val: 0.9465
Classifier:1, Fold: 3, AUC_train: 0.9886, AUC_val: 0.9559, F1-score_train: 0.9873, F1-score_val: 0.9467
Classifier:1, Fold: 4, AUC_train: 0.9897, AUC_val: 0.9417, F1-score_train: 0.9883, F1-score_val: 0.9268
Classifier:1, Fold: 5, AUC_train: 0.9887, AUC_val: 0.9542, F1-score_train: 0.9879, F1-score_val: 0.9453
Classifier:2, Fold: 1, AUC_train: 0.9926, AUC_val: 0.9489, F1-score_train: 0.9925, F1-score_val: 0.9377
Classifier:2, Fold: 2, AUC_train: 0.9929, AUC_val: 0.9577, F1-score_train: 0.9928, F1-score_val: 0.9482
Classifier:2, Fold: 3, AUC_train: 0.9923, AUC_val: 0.9562, F1-score_train: 0.9922, F1-score_val: 0.9478
Classifier:2, Fold: 4, AUC_train: 0.9929, AUC_val: 0.9569, F1-score_train: 0.9928, F1-score_val: 0.9470
Classifier:2, Fold: 5, AUC_train: 0.9903, AUC_val: 0.9529, F1-score_train: 0.9902, F1-score_val: 0.9439
Classifier:3, Fold: 1, AUC_train: 1.0000, AUC_val: 0.9377, F1-score_train: 1.0000, F1-score_val: 0.9253
Classifier:3, Fold: 2, AUC_train: 1.0000, AUC_val: 0.9478, F1-score_train: 1.0000, F1-score_val: 0.9354
Classifier:3, Fold: 3, AUC_train: 1.0000, AUC_val: 0.9554, F1-score_train: 1.0000, F1-score_val: 0.9465
Classifier:3, Fold: 4, AUC_train: 1.0000, AUC_val: 0.9395, F1-score_train: 1.0000, F1-score_val: 0.9269
Classifier:3, Fold: 5, AUC_train: 1.0000, AUC_val: 0.9495, F1-score_train: 1.0000, F1-score_val: 0.9399
[00:31:58] WARNING: C:/Users/Administrator/workspace/xgboost-win64_release_1.5.0/src/learner.cc:1115: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
Classifier:4, Fold: 1, AUC_train: 0.9997, AUC_val: 0.9452, F1-score_train: 0.9997, F1-score_val: 0.9326
[00:31:59] WARNING: C:/Users/Administrator/workspace/xgboost-win64_release_1.5.0/src/learner.cc:1115: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
Classifier:4, Fold: 2, AUC_train: 0.9997, AUC_val: 0.9530, F1-score_train: 0.9997, F1-score_val: 0.9429
[00:31:59] WARNING: C:/Users/Administrator/workspace/xgboost-win64_release_1.5.0/src/learner.cc:1115: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
Classifier:4, Fold: 3, AUC_train: 1.0000, AUC_val: 0.9538, F1-score_train: 1.0000, F1-score_val: 0.9441
[00:31:59] WARNING: C:/Users/Administrator/workspace/xgboost-win64_release_1.5.0/src/learner.cc:1115: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
Classifier:4, Fold: 4, AUC_train: 0.9994, AUC_val: 0.9480, F1-score_train: 0.9994, F1-score_val: 0.9345
[00:32:00] WARNING: C:/Users/Administrator/workspace/xgboost-win64_release_1.5.0/src/learner.cc:1115: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
Classifier:4, Fold: 5, AUC_train: 0.9994, AUC_val: 0.9537, F1-score_train: 0.9994, F1-score_val: 0.9452
Classifier:5, Fold: 1, AUC_train: 0.9980, AUC_val: 0.9411, F1-score_train: 0.9977, F1-score_val: 0.9293
Classifier:5, Fold: 2, AUC_train: 0.9992, AUC_val: 0.9587, F1-score_train: 0.9990, F1-score_val: 0.9485
Classifier:5, Fold: 3, AUC_train: 0.9990, AUC_val: 0.9556, F1-score_train: 0.9987, F1-score_val: 0.9456
Classifier:5, Fold: 4, AUC_train: 0.9997, AUC_val: 0.9464, F1-score_train: 0.9997, F1-score_val: 0.9321
Classifier:5, Fold: 5, AUC_train: 0.9987, AUC_val: 0.9447, F1-score_train: 0.9987, F1-score_val: 0.9326
Fold: 1, AUC_train: 0.9582, AUC_val: 0.9455, F1-score_train: 0.9491, F1-score_val: 0.9337
Fold: 2, AUC_train: 0.9538, AUC_val: 0.9495, F1-score_train: 0.9449, F1-score_val: 0.9398
Fold: 3, AUC_train: 0.9569, AUC_val: 0.9471, F1-score_train: 0.9477, F1-score_val: 0.9361
Fold: 4, AUC_train: 0.9546, AUC_val: 0.9556, F1-score_train: 0.9451, F1-score_val: 0.9456
Fold: 5, AUC_train: 0.9544, AUC_val: 0.9518, F1-score_train: 0.9463, F1-score_val: 0.9416
与LightGBM预测不同的样本数：  14
0     -1
8      1
23    -1
28    -1
33    -1
35    -1
47    -1
52    -1
60     1
64     1
74    -1
89    -1
796   -1
851   -1
Name: label, dtype: int64

4.8 归一化数据，pytorch神经网络

先归一化数据，统一量纲。从本节开始，使用基于距离的模型，不再使用树模型。

注意pytorch做二元分类有以下几种实现方式：

1. Linear 输出维度为1 + sigmoid + BCELoss。
2. Linear 输出维度为1 + BCEWithLogitsLoss。不需要加sigmoid或softmax函数，BCEWithLogitsLoss自带sigmoid作为激活函数。
3. Linear 输出维度为2 + 交叉熵（CrossEntropyLoss)。输出tensor的维度0对应第一个label（即0），维度1对应第二个label（即1）。注意使用交叉熵时，真实标签不能是onehot格式，必须为1维tensor，预测标签必须大于或等于2维，预测标签的每一个维度对应一个标签。

pytorch 中使用神经网络进行多分类时，网络的输出 prediction 是 one hot 格式，但计算 交叉熵损失函数时，loss = criterion(prediction, target) 的输入 target 不能是 one hot 格式，直接用数字来表示就行（4 表示 one hot 中的 0 0 0 1）。
所以，自己构建数据集，返回的 target 不需要是 one hot 格式。

Ref:
[1] Pytorch学习笔记(5)——交叉熵报错RuntimeError: 1D target tensor expected, multi-target not supported
[2] PyTorch二分类时BCELoss，CrossEntropyLoss，Sigmoid等的选择和使用
[3] Pytorch实现二分类器
[4] RuntimeError: multi-target not supported at

from sklearn.preprocessing import MinMaxScaler

# 归一化
scaler = MinMaxScaler()
X_train2 = scaler.fit_transform(X_train)
X_test2 = scaler.fit_transform(X_test)
Y_train2 = Y_train.to_numpy()
print('X_train.shape = ', X_train.shape)
print('X_train2.shape = ', X_train2.shape)
print('Y_train2.shape = ', Y_train2.shape)

X_train.shape =  (5070, 10)
X_train2.shape =  (5070, 10)
Y_train2.shape =  (5070,)

def Convert(x):
    # Conver the numeric values into categorical values.

    y = np.zeros((x.shape[0],))
    for i in range(len(x)):
        if x[i, 0] > x[i, 1]:
            y[i] = 0
        else:
            y[i] = 1

    return y

import torch
import torch.nn as nn
import torch.nn.functional as F

class NET(nn.Module):
    def __init__(self, input_dim:int, hidden:int, out_dim:int, activation='relu', dropout=0.2):
        super(NET, self).__init__()
        self.input_dim = input_dim
        self.hidden = hidden
        self.out_dim = out_dim
        self.activation = activation
        self.Dropout = dropout

        # 激活函数选择
        if self.activation == 'relu':
            mid_act = torch.nn.ReLU()
        elif self.activation == 'tanh':
            mid_act = torch.nn.Tanh()
        elif self.activation == 'sigmoid':
            mid_act = torch.nn.Sigmoid()
        elif self.activation == 'LeakyReLU':
            mid_act = torch.nn.LeakyReLU()
        elif self.activation == 'ELU':
            mid_act = torch.nn.ELU()
        elif self.activation == 'GELU':
            mid_act = torch.nn.GELU()

        self.model = nn.Sequential(
            nn.Linear(self.input_dim, self.hidden),
            mid_act,
            nn.Dropout(self.Dropout),
            nn.Linear(self.hidden, self.hidden),
            mid_act,
            nn.Dropout(self.Dropout),
            nn.Linear(hidden, self.out_dim)
        )

    def forward(self, x):
        out = self.model(x)

        return out

    def predict(self, x):
        # x = torch.tensor(x.to_numpy()) #针对datafram
        # x = x.to(torch.float32)
        x = torch.tensor(x).to(torch.float32)  #针对ndarray
        x = F.softmax(self.model(x))
        ans = []
        for t in x:
            if t[0] > t[1]:
                ans.append(0)
            else:
                ans.append(1)

        return np.array(ans)
        

import time
from torch.utils.data  import DataLoader,  TensorDataset
class NN_classifier():
    def __init__(self, model, crit, l_rate, batch_size, max_epochs, n_splits=5, verbose=True):
        super(NN_classifier, self).__init__()
        self.model = model # Neural network model, should be a nn.Module()
        self.l_rate = l_rate
        self.batch_size = batch_size
        self.max_epochs = max_epochs
        self.verbose = verbose
        self.n_splits = n_splits  # the value of k in k-fold validation
        self.crit = crit  # loss function
        self.device = 'cpu'

    def fit(self, X_train, Y_train, X_test):
        skf = StratifiedKFold(n_splits=self.n_splits)
        fold = 1
        pred_test = []
        for train_idx, val_idx in skf.split(X_train, Y_train):
            train_x = X_train[train_idx, :]
            train_y = Y_train[train_idx]
            val_x = X_train[val_idx, :]
            val_y = Y_train[val_idx]
            train_data = TensorDataset(train_x, train_y)
            train_dataloader = DataLoader(dataset=train_data, batch_size=self.batch_size, shuffle=True)
            valid_data = TensorDataset(val_x, val_y)
            validation_dataloader = DataLoader(dataset=valid_data, batch_size=self.batch_size, shuffle=False)
            
            model = self.model
            optimizer = torch.optim.Adam(model.parameters(), lr=self.l_rate)
            scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=50, gamma=0.9)  #动态学习率调整

            for epoch in range(self.max_epochs):
                start_time = time.time()
                loss_all = []
                #------------- Training -----------------
                model.train()
                for data in train_dataloader:
                    x, y = data
                    x = x.to(self.device)
                    y = y.to(self.device)
                    optimizer.zero_grad()

                    out = model(x)
                    loss = self.crit(out, y.long())
                    loss.requires_grad_(True) 
                    loss.backward()
                    optimizer.step()
                    loss_all.append(loss.item())
                    
                scheduler.step()
                end_time = time.time()
                cost_time = end_time - start_time
                train_loss = np.mean(np.array(loss_all))

                #------------- Validation -----------------
                model.eval()
                loss_all = []
                with torch.no_grad():
                    for data in validation_dataloader:
                        x, y = data
                        x = x.to(self.device)
                        y = y.to(self.device)
                        output = model(x)
                        loss = self.crit(output, y.long())
                        loss_all.append(loss.item())

                validation_loss = np.mean(np.array(loss_all))
                if self.verbose and (epoch+1) % 100 ==0:
                    print('Fold:{:d}, Epoch:{:d}, train_loss: {:.4f}, validation_loss: {:.4f}, cost_time: {:.2f}s'
                    .format(fold, epoch+1, train_loss, validation_loss, cost_time))

            #------------- Prediction -----------------
            pred = Convert(model(X_test).detach().numpy())
            pred_test.append(pred)

            pred_train = Convert(model(train_x).detach().numpy())
            pred_val = Convert(model(val_x).detach().numpy())

            auc_train = roc_auc_score(train_y, pred_train)
            auc_val = roc_auc_score(val_y, pred_val)
            f_score_train = f1_score(train_y, pred_train)
            f_score_val = f1_score(val_y, pred_val)
            print('Fold: %d, AUC_train: %.4f, AUC_val: %.4f, F1-score_train: %.4f, F1-score_val: %.4f'%(fold, 
                auc_train, auc_val, f_score_train, f_score_val))
            fold += 1

        pred_test = pd.DataFrame(pred_test).T
        print('pred_test.shape = ', pred_test.shape)
        # 将5次预测结果求平均值
        pred_test['average'] = pred_test.mean(axis=1)

        #因为竞赛需要你提交最后的预测判断，而模型给出的预测结果是概率，因此我们认为概率>0.5的即该患者有糖尿病，概率<=0.5的没有糖尿病
        pred_test['label'] = pred_test['average'].apply(lambda x:1 if x>0.5 else 0)

        ## 导出结果
        result=pd.read_csv('提交示例.csv')
        result['label']=pred_test['label']

        return result        

# k折交叉验证不断训练同一个模型，集成不同fold（即不同时刻）的模型的预测结果
hidden = 64
activation = 'relu'
# activation = 'tanh'
# crit = nn.MSELoss()
crit = nn.CrossEntropyLoss()
batch_size = 512*2
max_epochs = 500
l_rate = 1e-3
dropout = 0.1
n_splits = 5

# Convert to tensor
X_train_tensor = torch.from_numpy(X_train2).to(torch.float32)
X_test_tensor = torch.from_numpy(X_test2).to(torch.float32)
Y_train_tensor = torch.from_numpy(Y_train2).to(torch.float32)

model_NN = NET(X_train2.shape[1], hidden, out_dim=2, activation=activation)
classifier_NN = NN_classifier(model_NN, crit=crit, batch_size=batch_size, l_rate=l_rate, 
                    max_epochs=max_epochs, n_splits=n_splits)

result_SKFold_NN = NN_classifier.fit(classifier_NN,X_train_tensor, Y_train_tensor, X_test_tensor,)
c = result_LightGBM['label'] - result_SKFold_NN['label']

count = 0
for i in c:
    if i != 0:
        count += 1

print('与LightGBM预测不同的样本数： ', count)
print(c[c!=0])

Fold:1, Epoch:100, train_loss: 0.3503, validation_loss: 0.3196, cost_time: 0.04s
Fold:1, Epoch:200, train_loss: 0.2785, validation_loss: 0.2505, cost_time: 0.04s
Fold:1, Epoch:300, train_loss: 0.2332, validation_loss: 0.2257, cost_time: 0.04s
Fold:1, Epoch:400, train_loss: 0.2098, validation_loss: 0.2130, cost_time: 0.04s
Fold:1, Epoch:500, train_loss: 0.2006, validation_loss: 0.2066, cost_time: 0.04s
Fold: 1, AUC_train: 0.9363, AUC_val: 0.9133, F1-score_train: 0.9254, F1-score_val: 0.8956
Fold:2, Epoch:100, train_loss: 0.1813, validation_loss: 0.1555, cost_time: 0.04s
Fold:2, Epoch:200, train_loss: 0.1671, validation_loss: 0.1449, cost_time: 0.04s
Fold:2, Epoch:300, train_loss: 0.1540, validation_loss: 0.1444, cost_time: 0.04s
Fold:2, Epoch:400, train_loss: 0.1513, validation_loss: 0.1384, cost_time: 0.04s
Fold:2, Epoch:500, train_loss: 0.1426, validation_loss: 0.1379, cost_time: 0.04s
Fold: 2, AUC_train: 0.9496, AUC_val: 0.9395, F1-score_train: 0.9401, F1-score_val: 0.9269
Fold:3, Epoch:100, train_loss: 0.1356, validation_loss: 0.1419, cost_time: 0.04s
Fold:3, Epoch:200, train_loss: 0.1219, validation_loss: 0.1384, cost_time: 0.04s
Fold:3, Epoch:300, train_loss: 0.1206, validation_loss: 0.1357, cost_time: 0.04s
Fold:3, Epoch:400, train_loss: 0.1152, validation_loss: 0.1400, cost_time: 0.04s
Fold:3, Epoch:500, train_loss: 0.1113, validation_loss: 0.1395, cost_time: 0.04s
Fold: 3, AUC_train: 0.9625, AUC_val: 0.9550, F1-score_train: 0.9535, F1-score_val: 0.9434
Fold:4, Epoch:100, train_loss: 0.1075, validation_loss: 0.1185, cost_time: 0.04s
Fold:4, Epoch:200, train_loss: 0.1155, validation_loss: 0.1250, cost_time: 0.04s
Fold:4, Epoch:300, train_loss: 0.1081, validation_loss: 0.1238, cost_time: 0.04s
Fold:4, Epoch:400, train_loss: 0.1056, validation_loss: 0.1283, cost_time: 0.04s
Fold:4, Epoch:500, train_loss: 0.0957, validation_loss: 0.1289, cost_time: 0.04s
Fold: 4, AUC_train: 0.9702, AUC_val: 0.9518, F1-score_train: 0.9629, F1-score_val: 0.9386
Fold:5, Epoch:100, train_loss: 0.1064, validation_loss: 0.0951, cost_time: 0.04s
Fold:5, Epoch:200, train_loss: 0.0983, validation_loss: 0.0978, cost_time: 0.04s
Fold:5, Epoch:300, train_loss: 0.1028, validation_loss: 0.1055, cost_time: 0.05s
Fold:5, Epoch:400, train_loss: 0.0954, validation_loss: 0.1065, cost_time: 0.04s
Fold:5, Epoch:500, train_loss: 0.0935, validation_loss: 0.1073, cost_time: 0.04s
Fold: 5, AUC_train: 0.9709, AUC_val: 0.9547, F1-score_train: 0.9647, F1-score_val: 0.9455
pred_test.shape =  (1000, 5)
与LightGBM预测不同的样本数：  423
0     -1
2     -1
4     -1
8      1
16    -1
      ..
985   -1
987   -1
994   -1
995   -1
999   -1
Name: label, Length: 423, dtype: int64

k折交叉验证不断训练同一个模型，虽然模型最终的表现结果还可以（F1-score_val上去了），但集成各fold（各时期）的模型的结果表现依然糟糕，与baseline —— lightGBM相差甚远，都不用提交就知道分数会很低（0.63左右）了。

说明这网络模型不行啊！

可能原因：

• 集成时被早期表现较差的模型所拖累；
• 模型本身对表格数据拟合能力不够；
• 模型过拟合；

# 不用k折交叉验证，一个模型用到底
hidden = 64
activation = 'tanh'
# activation = 'tanh'
crit = nn.CrossEntropyLoss()
batch_size = 128
max_epochs = 2000
l_rate = 5e-3
dropout = 0.1
n_splits = 5

model_NN2 = NET(X_train2.shape[1], hidden, out_dim=2, activation=activation)
classifier_NN2 = NN(model_NN2, crit=crit, batch_size=batch_size, l_rate=l_rate, 
                    max_epochs=max_epochs)

# result_SKFold_NN = SKFold(pd.DataFrame(X_train2), pd.DataFrame(Y_train2),
#                     pd.DataFrame(X_test2), classifier_NN2, n_splits=5)

classifier_NN2.fit(X_train2, Y_train2)
result_NN = classifier_NN2.predict(X_test2)

# c = result_LightGBM['label'] - result_SKFold_NN['label']
c = result_LightGBM['label'] - result_NN

count = 0
for i in c:
    if i != 0:
        count += 1

print('与LightGBM预测不同的样本数： ', count)
print(c[c!=0])

Epoch:100, train_loss: 0.2839, cost_time: 0.09s
Epoch:200, train_loss: 0.2275, cost_time: 0.09s
Epoch:300, train_loss: 0.2072, cost_time: 0.09s
Epoch:400, train_loss: 0.2097, cost_time: 0.09s
Epoch:500, train_loss: 0.1854, cost_time: 0.09s
Epoch:600, train_loss: 0.1906, cost_time: 0.09s
Epoch:700, train_loss: 0.1772, cost_time: 0.09s
Epoch:800, train_loss: 0.1749, cost_time: 0.09s
Epoch:900, train_loss: 0.1726, cost_time: 0.08s
Epoch:1000, train_loss: 0.1707, cost_time: 0.08s
Epoch:1100, train_loss: 0.1692, cost_time: 0.09s
Epoch:1200, train_loss: 0.1575, cost_time: 0.09s
Epoch:1300, train_loss: 0.1615, cost_time: 0.09s
Epoch:1400, train_loss: 0.1548, cost_time: 0.09s
Epoch:1500, train_loss: 0.1611, cost_time: 0.09s
Epoch:1600, train_loss: 0.1617, cost_time: 0.09s
Epoch:1700, train_loss: 0.1535, cost_time: 0.09s
Epoch:1800, train_loss: 0.1650, cost_time: 0.09s
Epoch:1900, train_loss: 0.1609, cost_time: 0.08s
Epoch:2000, train_loss: 0.1631, cost_time: 0.09s
与LightGBM预测不同的样本数：  424
2     -1.0
4     -1.0
6     -1.0
7     -1.0
8      1.0
      ... 
986    1.0
988   -1.0
995   -1.0
997   -1.0
999   -1.0
Name: label, Length: 424, dtype: float64

上面这个结果说明这个神经网络模型学习遇到了瓶颈，很难再提升了。要么改模型，要么改训练方法（k折交叉训练重复训练同一个模型有提升，但提升有限，表现依然扑街），要么改数据。

4.9 SVM

表现不行。

Ref:
[1] Python3《机器学习实战》学习笔记（八）：支持向量机原理篇之手撕线性SVM

from sklearn.svm import SVC

model_SVM = SVC(C=10)  #C越大，对误分类的惩罚越大。

result_SKFold_SVM= SKFold(pd.DataFrame(X_train2), pd.DataFrame(Y_train2),
                    pd.DataFrame(X_test2), model_SVM, n_splits=5)

diff_SVM = evaluate(result_LightGBM, result_SKFold_SVM)

Fold: 1, AUC_train: 0.9001, AUC_val: 0.8614, F1-score_train: 0.8819, F1-score_val: 0.8311
Fold: 2, AUC_train: 0.8987, AUC_val: 0.8737, F1-score_train: 0.8797, F1-score_val: 0.8489
Fold: 3, AUC_train: 0.8942, AUC_val: 0.8916, F1-score_train: 0.8744, F1-score_val: 0.8711
Fold: 4, AUC_train: 0.8901, AUC_val: 0.8840, F1-score_train: 0.8688, F1-score_val: 0.8614
Fold: 5, AUC_train: 0.8945, AUC_val: 0.8836, F1-score_train: 0.8747, F1-score_val: 0.8602
与LightGBM预测不同的样本数：  271
2     -1
8      1
16    -1
23    -1
33    -1
      ..
983   -1
985   -1
993    1
995   -1
999   -1
Name: label, Length: 271, dtype: int64

4.10 sklearn神经网络

from sklearn.neural_network import MLPClassifier

model_MLP = MLPClassifier(hidden_layer_sizes=128, activation='relu')

result_SKFold_MLP = SKFold(pd.DataFrame(X_train2), pd.DataFrame(Y_train2),
                    pd.DataFrame(X_test2), model_MLP, n_splits=5)

c = result_LightGBM['label'] - result_SKFold_MLP['label']

count = 0
for i in c:
    if i != 0:
        count += 1

print('与LightGBM预测不同的样本数： ', count)
print(c[c!=0])

Fold: 1, AUC_train: 0.8780, AUC_val: 0.8394, F1-score_train: 0.8550, F1-score_val: 0.8045
Fold: 2, AUC_train: 0.8789, AUC_val: 0.8746, F1-score_train: 0.8558, F1-score_val: 0.8512
Fold: 3, AUC_train: 0.8838, AUC_val: 0.8828, F1-score_train: 0.8617, F1-score_val: 0.8599
Fold: 4, AUC_train: 0.8869, AUC_val: 0.8959, F1-score_train: 0.8637, F1-score_val: 0.8747
Fold: 5, AUC_train: 0.8867, AUC_val: 0.8820, F1-score_train: 0.8650, F1-score_val: 0.8579
与LightGBM预测不同的样本数：  238
2     -1
8      1
16    -1
23    -1
33    -1
      ..
979   -1
983   -1
993    1
995   -1
999   -1
Name: label, Length: 238, dtype: int64

神经网络在这个任务上的表现扑街了。。

5. 总结思考

这次比赛最好的分数为stacking的0.96577，再往上提升一点变得非常困难，继续提升一点分数需要耗费巨量时间和精力，投入产出比划不来，就没有继续去改进了。但从这次比赛也学到了许多，掌握了许多树模型的使用方法，以及特征工程的一点技巧。

在表格数据上，不得不说还是树模型表现更好，计算快，对算力的需求没有神经网络那么大，结果也非常棒。相反，神经网络在这个数据上表现明显不如树模型，或许是因为我采用的模型过于简单了。

透过现象看本质，这个比赛本质就是个简单的二分类问题，那么有没有一种可能，推荐系统里的DeepFM、DCN等网络模型也能用于这个问题呢？

另外，这个比赛我有点过于注重模型部分了，特征工程没有怎么去做。数据决定了你能达到的上限，模型只是帮你接近这个上限。

哈哈哈，有空再试了。欢迎各位大佬在评论区留言赐教，一起变得更强！

参考资料：
[1] Datawhale_如何打一个数据挖掘比赛V2.1
[2] 讯飞官方参考解析
[3] Kaggle上分技巧——单模K折交叉验证训练+多模型融合