CTR click through rate prediction practice project of advertising recommendation!

 Datawhale dried food  

author ： Fishman , Master of Wuhan University ,Datawhale member

And newspapers 、 The magazine 、 TV 、 Compared with these traditional media advertisements , New Internet advertising has natural advantages ： It can track 、 Study user preferences , And on this basis, carry out accurate advertising recommendation and marketing .

CTR（Click-Through-Rate） Click through rate , It is an important index to measure the effect of Internet advertising . This issue has become a hot topic in the research of major platforms in recent years . With the help of Huawei Global Campus AI Algorithm elite game question —— advertisement - Information flows across domains ctr forecast , Study this problem .

Practical background

Background of the contest

Advertising recommendation is mainly based on the historical exposure of users to advertisements 、 Click and other behaviors to model , If you only use advertising domain data , User behavior data is sparse , The type of behavior is relatively single . And the introduction of cross domain data from the same media , You can get the behavior data of the same advertising user in other domains , Deeply tap user interests , Enrich user behavior characteristics . Introduce advertising user behavior data of other media , It can also enrich the characteristics of users and advertisements .

Match task

This competition question is based on the advertisement log data , Basic user information and cross domain data to optimize advertising ctr Prediction accuracy . The target domain is the advertising domain , The source domain is the information flow recommendation domain , By obtaining the exposure of users in the information basin 、 Click on behavioral data such as information flow , Conduct user interest modeling , Help advertising domain ctr Accurate estimation .

Registration and data download

Registration address ：

https://developer.huawei.com/consumer/cn/activity/digixActivity/digixdetail/101655281685926449?ha_source=dw&ha_sourceId=89000243

Data download ：（ Students who have not participated in the competition refer to ）

https://xj15uxcopw.feishu.cn/docx/doxcnufyNTvUfpU57sRyydgyK6c

Practical ideas

This game is a classic hit rate prediction （CTR） Data mining competition , The task is to build a model , Predict whether the user clicks on the advertisement according to the user's test data . This is a typical dichotomous problem , The prediction output of the model is 0 or 1 （ Click on ：1, Did not click ：0）

Machine learning , About the classification task, we usually think of logical regression 、 Decision tree and other algorithms , In this practice code , We try to use logical regression to build our model . When we solve machine learning problems , Generally, the following process will be followed ：

Practice code

• Need memory ：1GB

• The elapsed time ：5 minute

# Install dependent Libraries   If it is windows System ,cmd Input in the command box pip install , Refer to the above environment configuration 
#!pip install sklearn
#!pip install pandas

#---------------------------------------------------
# Import library 
import pandas as pd

#---------------- Data exploration ----------------
#  Use only target domain user behavior data 
train_ads = pd.read_csv('./train/train_data_ads.csv',
    usecols=['log_id', 'label', 'user_id', 'age', 'gender', 'residence', 'device_name',
            'device_size', 'net_type', 'task_id', 'adv_id', 'creat_type_cd'])

test_ads = pd.read_csv('./test/test_data_ads.csv',
    usecols=['log_id', 'user_id', 'age', 'gender', 'residence', 'device_name',
    'device_size', 'net_type', 'task_id', 'adv_id', 'creat_type_cd'])
    
#---------------- Data set sampling ----------------
train_ads = pd.concat([
    train_ads[train_ads['label'] == 0].sample(70000),
    train_ads[train_ads['label'] == 1].sample(10000),
])

#---------------- model training ----------------
#  Load training logistic regression model 
from sklearn.linear_model import LogisticRegression
clf = LogisticRegression()
clf.fit(
    train_ads.drop(['log_id', 'label', 'user_id'], axis=1),
    train_ads['label']
)

#---------------- Results output ----------------
#  Model prediction and generate result files 
test_ads['pctr'] = clf.predict_proba(
    test_ads.drop(['log_id', 'user_id'], axis=1),
    )[:, 1]
test_ads[['log_id', 'pctr']].to_csv('submission.csv',index=None)

Practice improves

We have completed the cross domain advertising information flow ctr Estimated practical baseline Mission , Next, we can think from the following directions ：

• Continue to try different prediction models or feature engineering to improve the accuracy of model prediction

• Try strategies such as model fusion

• Check the advertising information flow across domains ctr Relevant data of prediction , Get other model building methods

Take part in the internal test

This paper is about Datawhale Project practice 2.0 course , If you are also a student , It's still in the introductory stage , You can enter the internal test learning group , We optimize the tutorial together in the learning feedback .

Sorting is not easy to , spot Fabulous Three even ↓