Research On Attention-Based Advertising Click-Through Rate Prediction Model

Online advertising is an extremely valuable business and is usually a highly automated system that auctions ad space for various publishers’ web pages that users visit.The auctioned ad space entails ranking eligible ads based on advertisers’ bids and ad click-through rates.Click-through rate prediction is analyzed by a platform that collects a range of data from users and then predicts the probability that they will click on an ad.In a platform with a large number of users,even a very small improvement in prediction accuracy may provide advertisers with significant revenue.In this paper,we propose a new solution after an in-depth study of new proposed techniques in various fields in recent years,and also design a new model,ResExInt.The main work of this paper is divided into the following aspects:(1)In this paper,we propose a feature interaction layer called ResExInteraction Layer,which uses two memory components from the external space to learn the features of all samples in the integer data set.In the selfattention mechanism,the self-attention focuses only on the information in its own sample features and ignores the association in different samples.In this paper’s method,the multi-head mechanism can capture different information in different subspaces,and then get the refined attention graph by two memory components.This method not only focuses on the association of different features in the same sample,but also trains the model to learn the features of different samples by sharing the information in the whole dataset through two memory components.(2)In order to better model feature interactions while exploiting the potential of the multi-head attention mechanism,this paper applies residual connections not only to different feature interaction layers,but also to different attention heads.The introduced residual connections improve the prediction results and speed up the convergence of the model during the training process with almost no increase in training time.Also,to further improve the performance,the SENet network is introduced to handle highdimensional features,allowing the model to focus more on information features and reducing the burden of subsequent interaction layer feature learning.(3)Experiments were conducted on three general-purpose datasets,and the experimental results showed that the model proposed in this paper achieved better results in different prediction tasks,which proved the stability and feasibility of the method proposed in this paper.Second,at the end of the experiments,ablation experiments were also conducted to analyze how much each part of the model proposed in this paper plays a role in the actual operation,and the experiments proved that each part plays a role in the performance of the model to different degrees.Overall,the results obtained from this experiment demonstrate that the model proposed in this paper is effective and feasible.