Research On Online Advertisement Click-through Rate Prediction Based On FM Deep Learning Model

In computational advertising,it is a very important machine learning problem to predict whether a user will click on an advertisement.Improve the accuracy of ad clickthrough rate predictions,providing personalized recommendations to users.On the one hand,it can solve the problem of how to efficiently obtain interested information under the condition of "information overload" and tap the potential needs of users.On the other hand,it can increase the revenue of the platform and advertisers.In the era of big data,each Internet platform needs to make effective use of a large amount of historical user behavior data accumulated on the platform to achieve accurate click-through rate prediction of advertisements.However,when using massive amounts of data to predict the click-through rate of advertisements,there are many problems.Click-through rate prediction has become a very important research topic.Based on domestic and foreign literature research,this article uses the massive advertising click-through rate data provided by the Huawei platform in 2020 and builds an online advertising click-through rate prediction model based on the GBDT+Deep FM model to estimate the probability of users clicking on ads.We verify the feasibility of the model from both theoretical and practical aspects.This paper mainly focuses on the following four aspects:Firstly,in the data processing part,based on the online advertising click-through rate data set in this paper,data exploration and visualization are carried out,data missing and coding are processed for different characteristic variables.In feature engineering,a reasonable feature engineering framework is constructed based on the understanding of the actual business background,including 30 basic statistical features,15 exposure features and 22 historical click-through rate features.A total of 102 features are used as feature variables of the online advertising click-through rate prediction model.Secondly,in view of the current problems of advertising click-through rate prediction and the high-dimensional sparse characteristics of click-through rate data sets,this paper proposes the GBDT+Deep FM model for online advertising clickthrough rate prediction,and introduces the principle and innovation of the model,which can satisfy both memory and generalized demand.Thirdly,in the part of prediction results,this paper sets different super parameters to compare and analyze the influence of different values of super parameters on the click-through rate prediction results of the model,and then obtains the optimal super parameters.By comparing the model prediction results of feature engineering,the feature engineering in this paper plays an important role in improving the accuracy of online advertising click-through rate prediction.In addition,we dig deeper into business data and use the GBDT model to filter out the top ten important features that affect online advertising click-through rate prediction.Finally,in the model comparison part,the GBDT+ Deep FM model is compared with the traditional machine learning model(LR,FM,GBDT,GBDT+LR)and the deep learning model(FNN,NFM,Wide&Deep,Deep FM)for the click rate prediction effect.The experimental results show that the AUC value of GBDT+Deep FM model prediction results is 0.7377,and the Log Loss value is 0.3123.Compared with other models,the GBDT+Deep FM model can increase the AUC value by 0.04%-7.5%,which can effectively improve the accuracy of online advertising click-through rate prediction.