| With the continuous development of Internet technology,many traditional industries have gradually shifted from offline to online.As an important part of the Internet,online advertising is increasingly appearing in people's field of vision.How to find advertisements that meet the needs of users and bring revenue to advertising platforms and advertisers through given user behavior data is the core problem for online advertising.To achieve accurate matching,the key is to accurately predict the users' click-through rate or conversion rate,which quantifies the effectiveness of ad serving.In addition,compared with click-through rate,the conversion rate is closer to the terminal of the advertising marketing chain,which means that it has stronger practical significance.There are various types of online advertising that cannot be covered by only one conversion rate prediction method.Therefore,this thesis selects the search advertising of the e-commerce platform as the research object.The e-commerce platform,which is a system that integrates an advertising platform and many advertisers,has a relatively complex structure and is easily affected by various factors.When shopping festivals such as "Double Eleven" and "618" arrive,a series of promotional activities of the platform and merchants will lead to drastic changes in user traffic,which means that there is significant difference between data distribution of the daily period and that of the shopping festival.The model learned from the data of the daily period is difficult to effectively predict the conversion rate of the shopping festival.Taking into account the above situation,this thesis aims to propose a method that can more accurately predict the conversion rate of shopping festival by reassessing and improving every aspects of the predicting method for daily period.The main work of this thesis include:(1)Summarize and compare commonly used machine learning algorithms,analyze the applicability of each algorithm to online advertising conversion scenarios,make certain optimizations,which provide a theoretical basis for multi-model ensemble.(2)Analyze and process the data related to the conversion of search advertising on the e-commerce platform during the shopping festival.As far as feature encoding is concerned,this thesis proposes a layered encoding method based on the principle of information entropy,which divides the encoding interval by quantization.In this method,the available high cardinality features are processed in combination with Mean Encoding and Word2 vec Encoding.(3)Based on the data analysis results,this thesis designs a new prediction method.Firstly,a data set division method is designed for shopping festivals.Secondly,corresponding feature engineering is constructed,which focuses on the conversion rate features.Finally,this thesis proposes a weighted average ensemble model,in which every single model uses a kind of gradient boosting tree algorithm to process input data so that the ensemble model only needs to build feature engineering once,which optimizes time efficiency.(4)With the evaluation indicators used to measure errors,the thesis compares and evaluates the prediction results after each modification of the new method,and proves the effectiveness and necessity of these modifications.The experimental results prove that the proposed conversion rate prediction method has high prediction accuracy,can effectively process different types of features,improve the utilization of data information,and more accurately predict the users' purchase intention than current methods.It is a method suitable for predicting conversion rate of search advertising on e-commerce platforms during the shopping festival. |
PDF Full Text Request