腾讯2018广告算法大赛思路解析 100行代码带你上0.73

比赛报名地址：algo.qq.com/person/mobile/landingPage?from=dsbryan

由于本次比赛我属于内部员工不得参赛，所以我尽量写一些思路解析，为大家提供一个baseline

github:

https://github.com/YouChouNoBB/2018-tencent-ad-competition-baseline

1.首先处理4个G的用户特征

    因为数据太大，而且不是能直接pandas读取的格式，所以需要做格式转换，用dict的方式来初始化DataFrame

2.拼接用户特征，广告特征

   训练数据中负样本的标签给的是-1，需要先转成0，预测数据的标签置为-1，方便合并后区分数据集。将缺失值填充为 '-1' ，为什么不是数值的-1呢？因为在LabelEncoder的时候需要对数据排序，同时存在string和int类型是无法比较的。所以需要填充为string类型的 ‘-1’。

3.将单取值的离散特征使用稀疏方式one-hot

   为什么要先将数据划分为训练集和测试集呢，因为稀疏的数据是无法分片的，所以只能先划分数据，分别拼接稀疏特征。如果使用pd.get_dummy()来获取onehot特征，生成的数据是可以用来分片的，但是稠密存储是个致命弱点。

    github上很多人问我train_x=train[['creativeSize']] 这句是什么意思，其实creativeSize这个特征是数值特征，不需要进行特别的处理，如果想处理的话可以考虑pd.cut来分段离散化。另一个原因是把这个特征拿出来构造一个新的DataFrame，方便和后面生成的稀疏特征进行拼接。所以使用的是[[]]来取值获得一个DataFrame，而不是[]取值来或者一个Seris

4.将多取值的离散特征使用稀疏方式向量化

   这个操作估计很多同学之前没有见过，一般出现在自然语言处理中，计算TF-IDF，LDA等时候使用，但是同样可以用来生成一个稀疏向量，作为新的特征，同时可以一个特征生成多个特征，比单独的处理更加方便。

    如果65行报错：empty vocabulary; perhaps the documents only contain stop words  ，初始化可以使用CountVectorizer(token_pattern='(?u)\\b\\w+\\b')

5.线下测试

    使用train_test_split划分数据，这行注释掉了。

6.线上提交

    线上预测的时候，模型训练中early_stopping_rounds 这个参数没什么用，参数n_estimatorsxu需要根据线下测试来重新指定。我看到有些同学设置为10000取得了0.74的成绩。。。。

# coding=utf-8
# @author:bryan
# blog: https://blog.csdn.net/bryan__
# github: https://github.com/YouChouNoBB/2018-tencent-ad-competition-baseline

import pandas as pd
import lightgbm as lgb
from sklearn.model_selection import train_test_split
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.preprocessing import OneHotEncoder,LabelEncoder
from scipy import sparse
import os

ad_feature=pd.read_csv('../data/adFeature.csv')
if os.path.exists('../data/userFeature.csv'):
    user_feature=pd.read_csv('../data/userFeature.csv')
else:
    userFeature_data = []
    with open('../data/userFeature.data', 'r') as f:
        cnt = 0
        for i, line in enumerate(f):
            line = line.strip().split('|')
            userFeature_dict = {}
            for each in line:
                each_list = each.split(' ')
                userFeature_dict[each_list[0]] = ' '.join(each_list[1:])
            userFeature_data.append(userFeature_dict)
            if i % 100000 == 0:
                print(i)
            if i % 1000000 == 0:
                user_feature = pd.DataFrame(userFeature_data)
                user_feature.to_csv('../data/userFeature_' + str(cnt) + '.csv', index=False)
                cnt += 1
                del userFeature_data, user_feature
                userFeature_data = []
        user_feature = pd.DataFrame(userFeature_data)
        user_feature.to_csv('../data/userFeature_' + str(cnt) + '.csv', index=False)
        del userFeature_data, user_feature
        user_feature = pd.concat([pd.read_csv('../data/userFeature_' + str(i) + '.csv') for i in range(cnt + 1)]).reset_index(drop=True)
        user_feature.to_csv('../data/userFeature.csv', index=False)
train=pd.read_csv('../data/train.csv')
predict=pd.read_csv('../data/test1.csv')
train.loc[train['label']==-1,'label']=0
predict['label']=-1
data=pd.concat([train,predict])
data=pd.merge(data,ad_feature,on='aid',how='left')
data=pd.merge(data,user_feature,on='uid',how='left')
data=data.fillna('-1')
one_hot_feature=['LBS','age','carrier','consumptionAbility','education','gender','house','os','ct','marriageStatus','advertiserId','campaignId', 'creativeId',
       'adCategoryId', 'productId', 'productType']
vector_feature=['appIdAction','appIdInstall','interest1','interest2','interest3','interest4','interest5','kw1','kw2','kw3','topic1','topic2','topic3']
for feature in one_hot_feature:
    try:
        data[feature] = LabelEncoder().fit_transform(data[feature].apply(int))
    except:
        data[feature] = LabelEncoder().fit_transform(data[feature])

train=data[data.label!=-1]
train_y=train.pop('label')
# train, test, train_y, test_y = train_test_split(train,train_y,test_size=0.2, random_state=2018)
test=data[data.label==-1]
res=test[['aid','uid']]
test=test.drop('label',axis=1)
enc = OneHotEncoder()
train_x=train[['creativeSize']]
test_x=test[['creativeSize']]

for feature in one_hot_feature:
    enc.fit(data[feature].values.reshape(-1, 1))
    train_a=enc.transform(train[feature].values.reshape(-1, 1))
    test_a = enc.transform(test[feature].values.reshape(-1, 1))
    train_x= sparse.hstack((train_x, train_a))
    test_x = sparse.hstack((test_x, test_a))
print('one-hot prepared !')

cv=CountVectorizer()
for feature in vector_feature:
    cv.fit(data[feature])
    train_a = cv.transform(train[feature])
    test_a = cv.transform(test[feature])
    train_x = sparse.hstack((train_x, train_a))
    test_x = sparse.hstack((test_x, test_a))
print('cv prepared !')

def LGB_test(train_x,train_y,test_x,test_y):
    from multiprocessing import cpu_count
    print("LGB test")
    clf = lgb.LGBMClassifier(
        boosting_type='gbdt', num_leaves=31, reg_alpha=0.0, reg_lambda=1,
        max_depth=-1, n_estimators=1000, objective='binary',
        subsample=0.7, colsample_bytree=0.7, subsample_freq=1,
        learning_rate=0.05, min_child_weight=50,random_state=2018,n_jobs=cpu_count()-1
    )
    clf.fit(train_x, train_y,eval_set=[(train_x, train_y),(test_x,test_y)],eval_metric='auc',early_stopping_rounds=100)
    # print(clf.feature_importances_)
    return clf,clf.best_score_[ 'valid_1']['auc']

def LGB_predict(train_x,train_y,test_x,res):
    print("LGB test")
    clf = lgb.LGBMClassifier(
        boosting_type='gbdt', num_leaves=31, reg_alpha=0.0, reg_lambda=1,
        max_depth=-1, n_estimators=1500, objective='binary',
        subsample=0.7, colsample_bytree=0.7, subsample_freq=1,
        learning_rate=0.05, min_child_weight=50, random_state=2018, n_jobs=100
    )
    clf.fit(train_x, train_y, eval_set=[(train_x, train_y)], eval_metric='auc',early_stopping_rounds=100)
    res['score'] = clf.predict_proba(test_x)[:,1]
    res['score'] = res['score'].apply(lambda x: float('%.6f' % x))
    res.to_csv('../data/submission.csv', index=False)
    os.system('zip baseline.zip ../data/submission.csv')
    return clf

model=LGB_predict(train_x,train_y,test_x,res)