Advertisement Conversion Rate Estimation Model Based On Improved Wide&Deep

With the popularity and rapid development of the Internet,Internet advertising is increasingly able to reach everyone through various channels.With huge traffic,Internet advertising has become the main means of profit for most Internet companies.At the same time,with the increase in computing power and the development of big data technology,deep learning,artificial intelligence and other technologies have made great progress,which has made major Internet giants and scientific research institutions have applied deep learning technology to the field of computational advertising.In the field of computational advertising,advertising conversion rate is the core issue of research.For an online advertisement,more clicks and more conversions are what advertisers and advertising platform want.The advertising conversion rate estimation model is dedicated to predicting the probability of an online advertisement being activated in the context after being clicked,that is,converting,so that it can be used in the ranking calculation of advertisement recommendations.Computational advertising around 2010 when computing power is scarce,many Internet companies used fast and easy-to-interpret logistic regression predictions.In the past ten years when the dimensionality of data has increased substantially,logistic regression has not been sufficient for adequate training and learning of high-dimensional sparse data.In 2016,Google pioneered the Wide&Deep model,which created a dual-grid structure for parallel training of linear and non-linear structures.Under the background of the time,it was already the model that was able to fully extract the cross-features in the data,and then it was used in the calculation of advertisements.All kinds of deep learning models appearing in the field follow the structural ideas of this model.This paper proposes an advertisement conversion rate prediction model based on improved Wide&Deep,which follows the structure of Wide&Deep dual-grid parallel training,drawing lessons from the construction ideas of NFM and DeepFM,and integrates field-aware factorization machines into it.This model can effectively learn high-level combined features,and at the same time can well cope with the challenge of massive sparse data.Experiments show that the model proposed in this paper has higher accuracy and usability than Wide&Deep and its derivative NFM and DeepFM,and can solve the actual problem of advertising conversion rate estimation.From the perspective of exploring the landing scenarios of the field-aware factorization machines and the characteristics in practical applications,or improving the accuracy of advertising conversion rate estimation,the research and exploration in this article are of practical significance.