【Python学习系列十七】基于scikit-learn库逻辑回归训练模型(delta比赛代码2)

机器学习任务流程：学习任务定义->数学建模->训练样本采样->特征分析和抽取->算法设计和代码->模型训练和优化(性能评估和度量)->泛化能力评估(重采样和重建模）；

算法思路：应用半监督学习思路，先用训练集训练出一个模型，然后用模型给预测集打标签，之后将打上标签的预测集也加入到训练集中用模型再训练，用f1-scror作为性能评估的依据。这个代码和之前比，主要是增加model.predict_proba()函数返回正例概率，自己设置阈值来选择正例样本。代码如下：

# -*- coding: utf-8 -*-

import pandas as pd
import time
from sklearn import metrics
from sklearn.linear_model import LogisticRegression
from sklearn import preprocessing
#from sklearn.tree import DecisionTreeClassifier

def main():
#省份和地市映射
data = {"province":['河北省', '山西省', '内蒙古自治区', '辽宁省', '吉林省', '黑龙江省', '江苏省', '浙江省', '安徽省', '福建省', '江西省', '山东省', '河南省', '湖北省', '湖南省', '广东省', '广西壮族自治区', '海南省', '四川省', '贵州省', '云南省', '西藏自治区', '陕西省', '甘肃省', '青海省', '宁夏回族自治区', '新疆维吾尔自治区', '北京市', '天津市', '上海市', '重庆市'],
"pro_code":[13,14,15,21,22,23,32,33,34,35,36,37,41,42,43,44,45,46,51,52,53,54,61,62,63,64,65,11,12,31,50]}
province = pd.DataFrame(data, columns = ["province", "pro_code"])
citydata=pd.read_csv(r"D:\city.csv")#加载地市映射表
#加载带标记数据
label_ds=pd.read_csv(r"D:\label.csv")
label_ds = pd.merge(label_ds, province, how = "left", on = "province")
label_ds = pd.merge(label_ds, citydata, how = "left", on = "city")
label_df = pd.DataFrame(label_ds[['denomination','min_amount','pro_code','age','sex','account_age','txn_count','use_nums',\
'txn_min_amount','txn_amount_mean','avg_discount','voucher_num','avg_txn_amt',\
'use_ratio','voucher_ratio','batch_no','voucher_no','city_id','label']])
label_df["denomination"] = label_df["denomination"].astype("int")
label_df["min_amount"] = label_df["min_amount"].astype("int")
label_df["pro_code"] = label_df["pro_code"].astype("int")
label_df["age"] = label_df["age"].astype("int")
label_df["sex"] = label_df["sex"].astype("int")
label_df["account_age"] = label_df["account_age"].astype("int")
label_df["txn_count"] = label_df["txn_count"].astype("int")
label_df["use_nums"] = label_df["use_nums"].astype("int")
label_df["txn_min_amount"] = label_df["txn_min_amount"].astype("int")
label_df["txn_amount_mean"] = label_df["txn_amount_mean"].astype("int")
label_df["avg_discount"] = label_df["avg_discount"].astype("int")
label_df["voucher_num"] = label_df["voucher_num"].astype("int")
label_df["avg_txn_amt"] = label_df["avg_txn_amt"].astype("int")
label_df["use_ratio"] = label_df["use_ratio"].astype("float")
label_df["voucher_ratio"] = label_df["voucher_ratio"].astype("float")
label_df["batch_no"] = label_df["batch_no"].astype("int")
label_df["voucher_no"] = label_df["voucher_no"].astype("str")
label_df["city_id"] = label_df["city_id"].astype("int")
label_df["label"] = label_df["label"].astype("int")
#加载未标记数据
unlabel_ds=pd.read_csv(r"D:\unlabel.csv")
unlabel_ds = pd.merge(unlabel_ds, province, how = "left", on = "province")
unlabel_ds = pd.merge(unlabel_ds, citydata, how = "left", on = "city")
unlabel_df = pd.DataFrame(unlabel_ds[['denomination','min_amount','pro_code','age','sex','account_age','txn_count','use_nums',\
'txn_min_amount','txn_amount_mean','avg_discount','voucher_num','avg_txn_amt',\
'use_ratio','voucher_ratio','batch_no','city_id','phone','voucher_no']])
unlabel_df["denomination"] = unlabel_df["denomination"].astype("int")
unlabel_df["min_amount"] = unlabel_df["min_amount"].astype("int")
unlabel_df["pro_code"] = unlabel_df["pro_code"].astype("int")
unlabel_df["age"] = unlabel_df["age"].astype("int")
unlabel_df["sex"] = unlabel_df["sex"].astype("int")
unlabel_df["account_age"] = unlabel_df["account_age"].astype("int")
unlabel_df["txn_count"] = unlabel_df["txn_count"].astype("int")
unlabel_df["use_nums"] = unlabel_df["use_nums"].astype("int")
unlabel_df["txn_min_amount"] = unlabel_df["txn_min_amount"].astype("int")
unlabel_df["txn_amount_mean"] = unlabel_df["txn_amount_mean"].astype("int")
unlabel_df["avg_discount"] = unlabel_df["avg_discount"].astype("int")
unlabel_df["voucher_num"] = unlabel_df["voucher_num"].astype("int")
unlabel_df["avg_txn_amt"] = unlabel_df["avg_txn_amt"].astype("int")
unlabel_df["use_ratio"] = unlabel_df["use_ratio"].astype("float")
unlabel_df["voucher_ratio"] = unlabel_df["voucher_ratio"].astype("float")
unlabel_df["batch_no"] = unlabel_df["batch_no"].astype("int")
unlabel_df["city_id"] = unlabel_df["city_id"].astype("int")
unlabel_df["phone"] = unlabel_df["phone"].astype("str")
unlabel_df["voucher_no"] = unlabel_df["voucher_no"].astype("str")

#模型训练和预测
f1_score_old=float(0)#f1-score
f1_score=float(0.3)#高于全部设置1的分数
outset=[]
flag=int(1)
label_df_cons=label_df#训练样本数不变
while (f1_score-f1_score_old)>0.0001 :#迭代收敛到f1-score不再提升
if flag==0 :#第一次训练排除样本数量带来的问题
f1_score_old=f1_score
#训练数据采样，80%训练，20%验证
print "总样本，有", label_df.shape[0], "行", label_df.shape[1], "列"
train_label_df=label_df#全量训练，ample(frac=0.8)
print "训练集，有", train_label_df.shape[0], "行", train_label_df.shape[1], "列"
test_label_df=label_df_cons.sample(frac=0.3) #用训练集来测试f1-score
print "验证集，有", test_label_df.shape[0], "行", test_label_df.shape[1], "列"

#模型训练
label_X = train_label_df[['pro_code','city_id','age','sex','account_age',\
'txn_count','txn_amount_mean','txn_min_amount']]
label_X = preprocessing.scale(label_X)#归一化
label_y = train_label_df['label']
model = LogisticRegression()
#if flag==0 :
#    model = LogisticRegression()#逻辑回归，第一次预训练
#else :
#    model = DecisionTreeClassifier()#决策树
model.fit(label_X, label_y)

if flag==0 :#模型验证，第一次训练不评分
expected = test_label_df['label']
predicted_X=test_label_df[['pro_code','city_id','age','sex','account_age',\
'txn_count','txn_amount_mean','txn_min_amount']]
predicted_X=preprocessing.scale(predicted_X)#归一化
predicted = model.predict(predicted_X)
f1_score = metrics.f1_score(expected, predicted) #模型评估
print f1_score
flag=int(0)
if f1_score_old<f1_score :#为未标记样本打上标记，然后加入训练集
unlabel_X=unlabel_df[['pro_code','city_id','age','sex','account_age',\
'txn_count','txn_amount_mean','txn_min_amount']]
unlabel_X_noScale=unlabel_X
unlabel_X=preprocessing.scale(unlabel_X)#归一化
unlabel_y=model.predict(unlabel_X)
out_y=pd.DataFrame(unlabel_y.reshape(-1,1),columns=['label'])
unlabel_X_new=unlabel_X_noScale.join(out_y,how='left')
label_df=pd.DataFrame()#原样本清空
label_df=label_df_cons.append(unlabel_X_new)#构成新的训练集
else : #迭代训练结束，输出结果
unlabel_X=unlabel_df[['pro_code','city_id','age','sex','account_age',\
'txn_count','txn_amount_mean','txn_min_amount']]
unlabel_info = unlabel_df[['phone','voucher_no']]
unlabel_X=preprocessing.scale(unlabel_X)#归一化
unlabel_y=model.predict_proba(unlabel_X)[:,1]#预测返回概率值，通过概率值阈值选择正例样本
out_y=pd.DataFrame(unlabel_y,columns=['prob']) #返回判定正例的比例
outset=unlabel_info.join(out_y,how='left')#输出结果
outset["label"] = outset.apply(lambda x:  0 if x["prob"] <0.57 else 1, axis = 1)
outset= outset[outset['label']==1]
outset=outset[['phone','voucher_no','label']]
outsetds=pd.DataFrame(outset)
outsetds.to_csv('D:\gd_delta.csv',index=False,header=None)#输出预测数据

#评价f1
#unlabel_X=pd.DataFrame(unlabel_X,columns=['pro_code','city_id','age','sex','account_age',\
#                  'txn_count','txn_amount_mean','txn_min_amount'])
#print unlabel_X.head(5)
#outset=unlabel_X.join(out_y,how='left')#输出结果
#outset["label"] = outset.apply(lambda x:  0 if x["prob"] <0.57 else 1, axis = 1)
#expected = outset['label']
#predicted_X=outset[['pro_code','city_id','age','sex','account_age',\
#                           'txn_count','txn_amount_mean','txn_min_amount']]
#predicted_X=preprocessing.scale(predicted_X)#归一化
#predicted = model.predict(predicted_X)
#f1_score = metrics.f1_score(expected, predicted) #模型评估
#print f1_score#0.855946148093
#退出循环
break

#执行
if __name__ == '__main__':
start = time.clock()
main()
end = time.clock()
print('finish all in %s' % str(end - start))

继续提升有三点：

1）可以尝试给预测集打标签用一个模型，迭代训练用另一个模型；

2）可以尝试抽取不同的特征来建模，其次对特征值做离散化处理；

3）可以尝试用部分特征来预训练，另一部分特征来做训练模型，可以降低过拟合问题；