Research And Implementation Of Advertising Click-through Rate Model Based On Attention Mechanism

With the rapid development of Internet technology,computational advertising has received extensive attention as an interdisciplinary subject.Computational advertising is mainly used to solve problems related to computing in online advertising,and aims to put advertisements to specific user groups.Advertising click-through rate estimation is a core issue in computational advertising.Good click-through rate estimation can increase revenue for advertisers,increase the flow of sponsor’s products,and bring users a better advertising experience.In the task of estimating the click-through rate of advertisements,the two problems of feature interactions and context data processing need to be solved.Wide&Deep and Deep FM models are usually used to solve the feature interactions problem,but they can’t construct any-order feature interactions.Youtube Net and DIN models are usually used to solve the problem of context data processing,but they don’t consider feature interactions and multiple context data processing.In order to solve these problems,advertising click-through rate models based on the attention mechanism are constructed.The main research work is as follows:In order to solve the feature interactions problem,deep feature-interaction network based on the attention mechanism is constructed,which uses attention mechanism and residual network to improve fully connected layer and factorization machine component in Deep FM,so that the network can construct any-order feature interactions.On this basis,deep context-interaction network based on pooling and attention mechanisms is constructed to process multiple context data,which uses pooling and attention mechanisms to process each context data,so that the network can deal with feature interactions and multiple context data problems simultaneously.In order to verify the validity of models,a series of comparative experiments are conducted on the Avito dataset,and AUC is selected as the evaluation index.The AUC values of Deep FM and deep feature-interaction network on the test set are 0.7432 and 0.7587 respectively.Among them,deep feature-interaction network is improved by 2.01% compared with Deep FM,indicating that it can perform feature interactions better.The AUC values of the deep context-interaction network based on pooling and attention mechanism on Avito are 0.7734 and 0.7938 respectively,which are improved by 1.84% and 4.61% compared to the deep feature-interaction network respectively,which shows that the deep context-interaction network can effectively deal with feature interactions and context data.