Improved Deep Learning Based Conversion Rate Estimation Of Mobile APP Advertising

In the mobile APP advertising conversion rate estimation scenario,user interactions have generated a large amount of data,and advertising CRE faces challenges such as high dimension,sparse,high order interactions among features,and difficulty in building user interest models.With the large scale of high-dimension heterogeneous data,for accurate CRE,it is greatly important to automatically extract the interactions among features and construct user interest models based on his/her behaviors.Deep learning has been proven to be feasible for such problems.Motivated by this,we here focuse on accurate CRE of mobile APP advertising based on improve deep learning.The main contents are as follows:(1)Feature Extraction based Conversion Rate Estimation with Improved Wide&Deep Model: Due to the high dimension,sparse,and high interactive,the accurate CRE faces great challenges.We propose an improved Wide&Deep model of fusing Field-aware Factorized Machine(FFM)and deep convolutional neural network(DCNN)to effectively and automatically obtain the low-order and high-order interactions of high-dimensional sparse features,so as to realize the automatic and efficient combination of features and improve the accuracy of mobile APP advertising conversion rate estimation.A feature combination algorithm based on width module FFM to extract the interaction relations of low-order features is presented for the embedding of sparse data.The extraction of the high-order interactive features based on a DCNN is further given by fusing the latent features obtained by the FFM.Finally,the interactive feature combinations obtained by width and depth modules are integrated for the CRE.The application of the proposed algorithm in predicting the conversion rate of Tencent's mobile APP advertisements demonstrates the effectiveness of the method in improving the prediction accuracy.(2)Converstion Rate Estimation based on User Multiple Dynamic Behaviors Extracted with Attention Enhanced Deep Learning: CRE is highly related with the users interests or preferences implying in their behaviors.Accordingly,the user behavior features induced CRE is further studied.Considering the dynamical and evolutionary feature of user behaviors,attention enhance deep learning is proposed to extract the user behaviors features,and then the CRE model is trained with such features.Attention articulated GRU is first constructed to obtain the single behavior series of a user,and such series are expressed as the dynamical and evolutionary features.Then,the self-attention is used to model the multiple behaviors features.The user multiple behavior series are merged into a vector and used to construct the CRE model.The proposed algorithm is applied to the CRE of Tencent's mobile APP advertisements,and the experiental results demonstrate its advantage in further improving the prediction accuracy.(3)Conversion Rate Estimation by Ensembling Interactive Features and User Multiple Behavoirs: Based on the contents(1)and(2),the ensembled CRE model is further concerned.First,the ensemble strategies and improvements of the single machine learning models,and then the framework of the features ensemble is presented.The averaging and stacking ensemble stratigies are applied to perform the task and the experimental results in mobile APP CRE demonstrate the algorithm's effectiveness.