Research Of Internet Advertising Click-through-rate Prediction Model

With the rapid development of Internet and e-commerce,computational advertising has received extensive attention as an emerging discipline that meets science.Computational advertising involves theories and techniques of many disciplines,including advertising,information retrieval,text analysis,statistical models,machine learning,and microeconomics.Computational advertising is designed to advertise to a specific audience,it has always been a hot spot in emerging Internet applications.Click-Through-Rate Prediction(CTR Prediction)is a key problem in the field of computing advertising.Click-through rate estimation is usually used to determine the probability of an ad being clicked by a user.Making predictions and identifying the ads that users are most likely to click on is one of the most important algorithms for advertising technology.The click-through rate of the ad is related to the order of the ad delivery and the click-to-charge.A good click-through rate estimation model can improve the platform revenue for the advertising platform,optimize the advertiser's product and budget,and bring a better advertising experience to the user.Internet advertising is mainly divided into search advertising and display advertising,and search advertising is the largest and fastest growing form of advertising.The advertising click rate is related to the ranking of advertising and click charging,and it plays an important role in the income of the entire search advertising.Taking search advertising as an example,this thesis uses the LightGBM framework to predict click-through rate.Compared with traditional machine learning algorithms,LightGBM has faster training speed,it also can handle large-scale data sets,and has low memory usage.Starting from massive data,the related theories and techniques of data preprocessing,feature selection and model fusion are systematically studied from this thesis.According to the common CTR prediction model,this paper mainly studies several models.The main improvement are feature extraction and model improvement,integrating advantages of several single models,fully exploiting linear and nonlinear features,then proposing an ensemble model.The model presents a better CTR prediction solution.The main work of the thesis is summarized as follows:(1)Analysis and processing of data.The data is analyzed and preprocessed on the existing datasets.The differences and distribution of the samples are analyzed from the perspective of statistical analysis,and the missing values are analyzed and processed.(2)Processing of features.Deleting irrelevant or redundant features,constructing statistical features based on the original data set,and performing One-Hot Encoder on the category features,testing the effects of statistical features under various models,and verifying the effectiveness of the constructed statistical features.(3)Processing of the model.This thesis implements the traditional Logistic Regression model and several popular machine learning algorithms(such as FM,FFM,RF)to predict the click-through rate,and based on the above model,proposes a ensemble model based on feature selection,the experimental results prove the validity of the model.