Research And Application Of Short Video Preference Model

Along with the make a spurt of progress of mobile communication technology and the popularization of various intelligent mobile terminals represented by smart phones and tablets,the mobile Internet advertising industry has developed rapidly,especially the mobile short video advertising.Short video preference rate(LTR)is an important factor to measure the advertising effect of mobile short video.Through the analysis and prediction of mobile short video advertising,users can not only browse their favorite videos,improve user experience,but also assist advertisers to use the budget reasonably,accurately deliver the advertising to the target population,and improve the website revenue of mobile media.In recent years,the research on the prediction of click through rate and preference rate has made great progress.At present,the most widely used prediction method in industry is logistic regression(LR).LR has the advantages of being simple and easy to implement largescale real-time parallel processing.However,the learning ability of linear model is limited,which can not capture the information carried by high-order features,thus limiting the prediction performance of LR.However,the prediction of short video user preference rate often faces many problems,such as large amount of data,many dimensions of data features,and many topics of data.How to obtain valuable high-level information from different themes,different dimensions of features and combinations of features is very important for the short video advertising platform to enhance the value of the platform.On account o the over problems,the main contents of this paper are as as below:(1)the first mock exam model of LR,GBDT,FM and XGBoost and FFM are studied.By comparing and analyzing the advantages and disadvantages,based on the idea of GBDT+LR fusion model proposed by Facebook,XGBoost+FFM fusion model is proposed to further enhance the overall nonlinear learning ability of the fusion model,and then enhance the prediction effect of short video user preference rate.(2)In order to make full use of the multi topic nature of short video advertising,a short video preference prediction model based on the topic model is designed and implemented.Using LDA to model the short video title to get the short video theme distribution,then using it to segment the original data set,using fusion model to model the data set under different themes,finally using the synthesis strategy to integrate the sub models to get the final prediction.So that the model can learn the influence of feature combination of different topics on the results,and then improve the learning ability of the model.(3)In this paper,the short video data of bytedance,an Internet company,is taken as the analysis object.The results show that on the basis of not significantly increasing the training time,the AUC performance indexes of the short video preference prediction model based on the fusion model are 1.2% The prediction models based on the theme model include 2.3% The promotion of.