Research On CTR Prediction Model Considering Feature Importance

The vigorous development of the Internet has led to the overall change of the society and the rapid development of social productivity.The combination of Internet and traditional industries has changed people's life style and has become an indispensable part of people's life.For Internet companies,advertising realization is one of the means on which companies rely for survival.Click-through rate estimation is an important part of advertising realization,and its effect directly affects the success rate of advertising delivery.The most common indicator of click-through rate estimation in the industry is AUC.It can directly reflect the effect of advertising click-through rate estimation: the greater the value of AUC,the better the advertising click-through rate estimation effect.Therefore,how to improve the value of AUC in the advertising click-through rate prediction model is a problem of general concern in academia and industry.Through summarizing the research results of the advertising click-through rate prediction model,this paper finds that the existing advertising click-through rate prediction model has the following problems.First,the existing click-through rate prediction models of deep learning,such as DCN,NFM,Deep FM,etc.,focus on capturing the interaction of features while ignoring the importance of features themselves.Secondly,with the increase of the network layer of deep learning model,the training difficulty of the model will increase,leading to poor training effect of the model.Some models,such as NFM and DCN,try to introduce the normalization method for improvement,but they are not deep enough.In this context,this paper studies the click-through rate prediction model of frontier deep learning,takes the factor of feature importance into consideration,proposes a new click-through rate prediction model based on feature importance,and applies the normalization method to this model.The main content of this paper includes the following aspects:(1)Conduct research and analysis on click-through rate prediction model and related technologies of normalization method,and elaborate on the existing click-through rate prediction model and relevant knowledge of ECA module on this basis.(2)As the existing click-through rate prediction models ignore the importance of features before feature interaction,this paper proposes a click-through rate prediction model EPBi NET based on feature importance.The model is composed of Embedding layer,ECANET network layer,Bi-Interaction layer,Product layer and DNN network.The model converts the features into low-dimensional and dense Embedding vectors through the Embedding layer,and then uses the ECANET network layer to dynamically learn the weight of the features and multiply the resulting weight with the original Embedding vector to get a new weighted vector.Then,features are interacted through the Bi-Interaction layer,Product layer and DNN network to capture higher-order features.On the two groups of open data sets,the EPBi NET model verified by comparative experiments is better than the experimental model in terms of AUC.(3)In view of the limited research on the application of the normalization method to the existing click-through rate prediction model,this paper explores the application of the normalization method to the Embedding and MLP of the click-through rate prediction model,and proposes a new normalization method,ABLN,in the click-through rate prediction.This method combines the characteristics of BN and LN,and selects the appropriate normalization method automatically in an adaptive way.Through a large number of experimental comparisons,it is found that the normalization method can improve the effectiveness of the original model when applied to the Embedding and MLP of the model.In addition,the AUC of the model can be greatly improved by the way that the ABLN method acts on the Embedding layer and the LN method acts on the MLP part.