Research On The Conversion Rate Of Mobile APP Advertising

With the widespread use of tablet computers and smart phones,built-in advertising in apps has gradually become an important form of mobile digital marketing.Firstly,the advertiser will provide the advertising and advertising-related historical information to the mobile APP advertising platform,and then the advertising platform through the analysis of advertising data and advertising strategy,finally find the right user for advertising display.It is very challenging for advertising platforms to develop effective and efficient recommendation methods.Generally,two important indicators should be considered in advertising recommendation algorithms: CTR(click through rate)and CVR(conversion rate).When recommending advertisements,many existing advertising platforms aim to maximize the number of clicks.However,mobile APP advertisements expect users to download the APP instead of just clicking.The action of downloading the APP actually converted the advertisement.The advertising platform improves the display value of advertisements by predicting the probability of the advertisements being converted by users and then arranging the advertising items to be displayed according to the CVR.If the predicted CVR is too high,advertisers will waste money on the budget.On the contrary,if the predicted CVR is too low,advertisers will miss the opportunity to show with high CVR.In fact,because so little advertising data is converted,predicting CVR is more challenging than predicting CTR.Moreover,most of the advertising information in real life is sparse and huge,so it is difficult to predict its conversion rate.However,many existing models can handle sparse high-dimensional data problems of advertising information,but there is a lack of research on the interaction between deeper features.Combining user information,advertising information and context information,this paper proposes a new hybrid model based on FFM(field-aware factorization machines).This model is a new hybrid model: Through LightGBM(light gradient boosting decision tree)to effective learning high-dimensional combination characteristic,at the same time using FFM model can deal with highly sparse and the discrete feature,increase the precision of the existing model predicted CVR.The main innovative work of the paper includes:(1)This paper constructed a new hybrid model,which uses the output of LightGBM model as part of the input of FFM model to realize the prediction of advertising conversion rate.(2)In order to extract the input features needed in training model,on the one hand,low-dimensional features are extracted from the original data,on the other hand,high-dimensional combination features are extracted from LightGBM model.In order to transform the high-dimensional combined features into the input form of FFM,a new feature transformation form is proposed,which realizes the fusion of different models.(3)In addition,the prediction performance of the model is optimized by experimental analysis from the conversion data and model parameters.In order to verify the effectiveness and generalizability of the hybrid model,it is compared with other models on different data sets.The experimental results show that the prediction accuracy of the hybrid model is the highest.